Symposium S.EL06—Photonic Materials for Information Processing and Computing

Semiconductor materials play a pivotal role in photonic technology. They can be used as tailored photon emitters and detectors. Combined with plasmonic structures and photonic design, their efficiency and functionality is further extended.
In this talk we show how compound semiconductor nanostructures such as nanowires and nananoscale membranes (nanofins) can be used for both single photon emission and detection [[1],[2]]. We will explore what kind of heterostructures are optimal for these applications. We will also elucidate how photon detection can be optimized by the nanowire geometry and the combination with plasmonic structures [2,[3]]. Finally, we will also explore materials with a lower production and environmental cost and how to fabricate them at a similar functionality than compound semiconductors. We will show how GeSn thin films and nanostructures have the potential to become the leading material in mid-infrared photodetection.
[1] M. Heiss et al Nature Mater. 12, 439 (2013)
[2] A. Dorodnyy et al IEEE J. Sel. Top. Quant. Electr. 24, 1-13 (2018)
[3] A. Casadei et al, Sci. Rep. 5, 7651 (2015)

Manipulation of single photons generated deterministically from pre-specified sources on-chip to create multiphoton entangled states remains a major goal to be reached in photonic quantum information
processing platforms. Notable advances have been made in single photon source performance and
demonstrations of photon interference and entanglement based largely upon finding suitable single photon
source from an otherwise random ensemble. The lack of scalable on-chip architectures stands as a major
obstacle to realizing compact quantum information technologies.
In this talk I will present our proposition [1,2] and continuing efforts [3,4] to realize such a
photonic chip built upon a unique class of spatially-ordered arrays of surface-curvature driven mesa-top
single quantum dots (MTSQDs) [1,3]. These AlGaAs/InGaAs MTSQDs exhibit spectral uniformity as
low as 1.8nm across 5×8 arrays, exhibit pairs of as-grown QDs having emission energy within 200μeV,
and single photon purity > 99%. Following a planarization overlayer growth, these buried single photon
sources (SPS) provide the essential platform for subsequent lithographic fabrication of emitted photon
light manipulating units (LMUs). The LMUs exploit either the well-developed photonic 2D crystal
approach or a new approach we introduced [1,2] based upon Mie-like resonances in interacting
subwavelength-sized dielectric building block (DBB) based metastructures. The in-plane single Mie-like
resonance of the metastructure (the LMU) provides all the needed basic photon manipulation functions to
create controlled on-chip interference and entanglement [4]: enhancement of the SPS emission rate
(Purcell enhancement), directional emission (local antenna), state-preserving propagation, path
bifurcation (beam-splitting), and beam combining. Findings of our combined theoretical analysis,
numerical simulations, and experimental studies of such SPS-LMU primitives — the essential building
unit for quantum optical circuits — will be presented. The emphasis is on the holistic approach to proofof-
principle demonstration of the needed on-chip multifunctional system with its built-in trade-offs rather
than the best of any individual light manipulation function.
The work is supported by the US Army Research Office (W911NF1910025) and the Air Force Office of
Scientific Research (FA9550-17-1-0353).
1. Jiefei Zhang et al., Jour. App. Phys. 120, 243103 (2016)
2. Swarnabha Chattaraj et al., Jour. Opt. Soc. America B. 33, 2414(2016)
3. Jiefei Zhang, et al., App. Phys. Lett 114, 071102 (2019)
4. Swarnabha Chattaraj et al., IEEE Jour. Quant. Elec. 56, 1 (2019)

Since the first proposal of the semiconductor quantum dot in 1982[1]. the quantum dots have been intensively investigated for both fundamental solid state physics and device applications. Advances of self-assembling crystal growth technology of quantum dots enabled realization of high performance semiconductor lasers[2] and quantum information devices such as single photon sources. Moreover, implementing a single quantum dot in an optical nanocavity provides a new platform for solid-state cavity quantum electrodynamics (cavity-QED)[3].
In this talk, we first discuss recent progress in growth and physics of GaN quantum dots for single photon sources (SPSs) at a ultraviolet wavelength, including a position controlled nanowire quantum dot with a giant biexciton binding energy over 60meV and high temperature operation up to 350 K[4,5]. Secondly, for implementation of SPSs onto silicon-based photonic integrated circuits, we also discuss a transfer printing method that integrates quantum dot SPSs into a silicon waveguide, demonstrating InAs/GaAs quantum dot SPSs with high waveguide coupling efficiencies, as well as the integration of two SPSs into a waveguide[6].

References
[1] Y. Arakawa and H. Sakaki, Appl. Phys. Lett. 40, 939 (1982)
[2] K. Nishi, K. Takemasa, M. Sugawara, and Y. Arakawa, IEEE J. Sel. Top. Quantum Electron., 23 16921013 (2017)
[3] M. Nomura, N. Kumagai, S. Iwamoto, Y. Ota, and Y. Arakawa, Nat. Phys. 6, 279 (2010)
[4] M. Holmes, K Choi, S. Kako, M. Arita, and Y. Arakawa, NanoLett., 14 982 (2014).
[5] M. Holmes, K Choi, S. Kako, M. Arita, and Y. Arakawa, ACS Photonics 3 543 (2016).
[6] R. Katsumi, Y. Ota, M. Kakuda, S. Iwamoto, and Y. Arakawa, Optica 10, 691 (2018).

Superconducting qubit quantum computers offer new capabilities for quantum information processing, but require < 50 mK environments and near complete isolation from outside noise sources. To enable quantum information to be exchanged with these systems outside of cryogenic conditions by means of higher energy photons, many approaches towards microwave-optical transduction are being explored by the research community. Three-wave mixing in linear electro-optic materials is a particularly appealing approach due to the inherent simplicity and elimination of excess noise sources that are introduced by intermediate states present in many alternative technologies. However, material microwave and optical loss rates have limited conversion efficiency below 50% in demonstrations to date. Here, I present IBM’s work towards integrating low loss SiGe/Si optical resonators with the low loss Nb-on-Si microwave resonators that have been developed for superconducting qubits. An effective linear electro-optic coefficient for coupling the microwave and optical fields by three wave mixing is induced by strongly biasing the microwave resonators to leverage the DC Kerr effect in Si and SiGe. For 1 MHz-rate transducers, the resulting microwave-optical coupling strength is calculated to be sufficient to reach the peak conversion efficiency condition. Demonstrating microwave-optical transduction in this platform is then dependent on achieving low defectivity fabrication of the proposed SiGe/Si optical waveguides. I will present a study of this materials problem and an analysis of the overall optimal transducer design.

This work was funded by LPS/ARO under CQTS program, contract number W911NF-18-1-0022.

Photonic quantum technologies rely on the fact that photons can be endowed with long coherence times, can travel long distances at light-speed, and interact only weakly with the environment, thereby constituting near-ideal carriers of quantum information. Single-photon qubits are a core resource for many quantum technologies currently being developed for systems that, based on quantum mechanical effects, have great potential to perform communication, signal processing, computation and measurements with significantly superior security, efficiency and sensitivity [1–3]. In such applications, an ideal quantum light source would emit a pure, high-rate stream of highly indistinguishable single-photons produced on-demand. Single-photon sources based on single quantum emitters – such as quantum dots, organic molecules, point-defects in solids, etc. - have a great potential to fulfill all such requirements. Among the various types of single quantum emitters available in the solid-state, self-assembled InAs/GaAs quantum dots constitute the most promising material system to date [4], having been used to demonstrate triggered single-photon emission with indistinguishability close to that of spontaneous pair sources, and orders or magnitude higher generation rates [5, 6]. Such quantum dots consist of nanometer-scale islands of InAs that are grown epitaxially in a GaAs host, and are operated at cryogenic temperatures. Importantly, the ability to produce quantum dots within high index contrast GaAs nanophotonic geometries enables strong control of crucial quantum dot emission characteristics, such as the excited state decay rate and outcoupling into select spatial modes [5–7]. The former characteristic translates into control of the single-photon emission rate and, to some degree, coherence. The latter translates into efficient funneling of the single-photons into a useful, output optical spatial mode – e.g., a low-divergence Gaussian beam [7, 8] or a single optical fiber mode [9] -, which has a strong impact on the overall source efficiency. This talk will describe our progress in developing high efficiency single-photon sources based on self-assembled InAs/GaAs quantum dots in nanophotonic structures. In particular, because the availability of mature chip-scale photonic integration techniques provides compelling advantages towards scalable photonic quantum systems, our recent results on deterministic fabrication of heterogeneous integrated photonic devices with single quantum dot based sources of on-chip waveguide-coupled single-photons [10, 11] will be covered.
References

[1] - J. L. O’Brien, A. Furusawa, and J. Vuckovic, Nature Photonics 3, 687 (2009).
[2] - M. D. Eisaman, J. Fan, A. Migdall, and S. V. Polyakov, Rev. of Sci. Instr. 82, 071 101 (2011).
[3] - N. Sangouard and H. Zbinden, J. Mod. Opt. 59, 1458 (2012).
[4] - I. Aharonovich, D. Englund, and M. Toth, Nature Photonics 10, 631 (2016).
[5] - N. Somaschi, et al., Nature Photonics 10, 340 (2016).
[6] - X. Ding, et al., Phys. Rev. Lett. 116, 020 401 (2016).
[7] - M. Davanco, et al., App. Phys. Lett. 99, 041 102 (2011).
[8] - L. Sapienza, et al., Nature Communications 6, 7833 (2015).
[9] - M. Davanco, et al., App. Phys. Lett. 99, 121 101 (2011).
[10] - M. Davanco,et al., Nature Communications 8, 889 (2017).
[11] - P. Schnauber et al., Nano Letters 19 7164 (2019).

Single-photon detectors are increasingly becoming an essential tool for a wide range of applications in physics, chemistry, biology, communications, medicine, and remote sensing. In the area of computing, efficient single photon detection is important for quantum information processing with photons as well as recently proposed work for large-scale superconducting optoelectronic networks for neuromorphic computing. Ideally, a single photon detector generates a measurable signal only when a single photon is absorbed. Furthermore, the ideal detector would have 100% detection efficiency, no false positive (dark counts), and transform-limited timing resolution. Recently, there has been tremendous progress in the development of superconducting nanowire single-photon detectors (SNSPDs or SSPDs) towards nearly ideal performance. The SNSPD is an alternative to conventional semiconductor avalanche photodiodes (APDs) especially for wavelengths in the near-infrared region of the spectrum. Since the first demonstration of single photon detection with a superconducting nanowire, there has been significant effort to package nanowire detectors into systems that could be used in real-world applications. I will review a few technological breakthroughs in SNSPD designs and performance, will briefly describe where advances in materials and material science could have immediate impact on the use of these detectors for computing.

Optically active spin defects in wide bandgap semiconductors continue to attract attention as qubits in the solid state. The ground state spin sublevels of these defects can store and process information while the atom-like optical transitions can initialize and read out the spin. Introducing these defect centers reliably into the host material is critical for future device integration. To this end, we explore chromium ion creation through implantation and annealing in silicon carbide (SiC) [1], a technologically mature and CMOS compatible wafer scale host crystal. Chromium in the 4+ charge state (Cr⁴⁺) has an S = 1 ground state and an S = 0 excited state with a near-infrared (1070 nm) strain insensitive optical transition that has a high Debye-Waller factor of 75%, making it interesting for quantum communication through fiber optic networks. We find that the created Cr⁴⁺ ensemble has an order of magnitude narrower linewidth of the optical hole (31 MHz) compared to as grown samples. We use the Λ-like level structure as a spin-photon interface and study the transition rates to optimize for high fidelity spin initialization and readout of 79%. At cryogenic temperatures (15 K), we measure long ensemble spin coherence times of T₂^* = 317 ns and a T₂ = 81 μs limited by the ensemble density and T₁ greater than 1 s. These results demonstrate Cr⁴⁺ in SiC as an attractive optically active spin-qubit for integration within quantum devices.
[1] B. Diler, et. al. arXiv:1909.08778
Funding Agencies: DOE

The neutral divacancy (VV0) in silicon carbide (SiC) exhibits robust spin coherence and a high-quality near-infrared spin-photon interface in a material compatible with mature fabrication techniques. Here, we make use of this scalable semiconductor host and design electronic devices to manipulate embedded isolated quantum systems. Specifically, we create and isolate single spin defects in a commercial p-i-n diode [1]. This simple integration enables engineering of the defect’s charge environment and drastically reduces electric field noise. Surprisingly, the use of electrical gating mitigates spectral diffusion and achieves near-lifetime limited optical linewidths. Furthermore, by exploiting field confinement of the junction, we show that the optical transitions can be gate-tuned by nearly a terahertz, revealing a spectral tuning range extending over 40,000 optical linewidths. This geometry may allow for spectral multiplexing of many quantum channels, and also provides a method for using electric fields combined with optical excitation for deterministic charge state control.

Applying gigahertz ac electric fields to SiC devices produces coherent interference in the form of Landau-Zener-Stückelberg fringes, arising from interactions between microwave and optical photons [2], even in the absence of a microwave resonator. We demonstrate lifetime-limited optical coherence and clock-like spin transitions with increased robustness against magnetic noise. Electrical driving of excited-state electron orbitals offers advantages over spin-based coupling and points towards new types of hybrid quantum systems. These results reveal new opportunities for electrical manipulation of spin-based quantum systems in scalable SiC electronic devices.

This work was done in collaboration with S. L. Bayliss, A. L. Crook, P. J. Mintun, S. J. Whiteley, G. Wolfowicz, H. Abe, A. Gali, V. Ivady, T. Ohshima, G. Thiering, P. Udvarhelyi.

[1] C. P. Anderson*, A. Bourassa*, K. C. Miao, G. Wolfowicz, P. J. Mintun, A. L. Crook, H. Abe, J. U. Hassan, N. T. Son, T. Oshima, D. D. Awschalom, “Electrical and optical control of single spins integrated in scalable semiconductor devices,” Science, accepted for publication (2019); arXiv: 1906.08328

[2] K. C. Miao, A. Bourassa, C. P. Anderson, S. J. Whiteley, A. L. Crook, S. L. Bayliss, G. Wolfowicz, G. Thiering, P. Udvarhelyi, V. Ivady, H. Abe, T. Ohshima, A. Gali, D. D. Awschalom, “Electrically driven optical interferometry with spins in silicon carbide,” Science Advances, in press (2019); arXiv: 1905.12780

In recent years, vacancy related defect centers in SiC have been identified as optically active, high potential candidates for the application in devices in the field of quantum information. Often such defect centers have been created by means of ion or electron beam implantation followed by subsequent annealing. As a result, these materials remain defective from a structural point of view with deteriorate electronic properties. In this work we present a process to generate and tailor vacancy defect centers in-situ during sublimation epitaxial growth of 3C-SiC with preserved electronic properties. (i) By adaption of the growth rate we were able to tailor the concentration of the carbon vacancy defect VcVsi, VcCsi and Vc at the discrete optical transition energies between 1.138 eV, 1.233 eV and 1.278 eV, respectively. (ii) Using extreme annealing conditions, we observed the generation of the nitrogen vacancy defect (NcVsi)- at an optical transition energy of 0.884 eV. In our presentation we will introduce the processing of the special 3C-SiC layers, identify their structural properties using Raman spectroscopy and discuss the optical active defect centers by means of low temperature photoluminescence measurements.

High-density and high-efficiency small light sources will have a wide range of applications including sensing, spectroscopy, quantum computing, and optical communications. In this regard, the integration of III-V materials on silicon has been pursued for decades to combine the best of both semiconductor families: Silicon for electronics and photonics components such as passive waveguides, wavelength filters, modulators, and photodetectors along with III-V semiconductors as direct bandgap efficient light emitters. Several approaches have been investigated relying overall on either bonding III-Vs to a silicon wafer or direct growth.
The main challenge to be overcome in direct growth stems from the significant lattice and thermal mismatch between III-Vs and Si that leads to crystal defects. We have developed a novel growth method named template-assisted selective epitaxy (TASE) to cointegrate seamlessly different materials.1 The high crystal quality of TASE-grown III-Vs resulted in state-of-the-art electronic devices.2 TASE-grown materials fill and take the shape of a pre-defined hollow oxide template. Overall benefits of TASE include low defect density, absence of foreign catalyst (Au-free), and precise control of crystal composition, position, shape, and size.
For integrated photonics TASE offers several advantages that are not achievable with other integration schemes. As all the fabrication steps involve only Si – no III-V processing – the template can hence be fabricated in a Si photonic foundry on an SOI wafer. This way, based on the mature Si technology, we can create III-V nanostructures featuring arbitrary shapes, e.g., vertical or horizontal nanowires, rings, microdisks, or photonic crystals. It also enables in-plane co-integration with Si-based components. Dense co-integration of different III-Vs was implemented on the same chip by sequential growth runs.3
We have recently demonstrated microdisk lasers grown by TASE which compared favorably to similar structures fabricated by the mature wafer bonding approach.4 However, the true benefit of TASE comes from enabling a seamless integration with silicon. In this talk we shall focus on these aspects, and how this enables novel device architectures including hybrid III-V/silicon photonic crystal cavities5 and photodetectors.6
This research was supported by the European Union H2020 ERC StG PLASMIC (Grant No. 678567), SiLAS (Grant No. 735008) and Swiss National Science Foundation SPILA (Grant No. CRSK-2_190806).
References
1. Schmid, H. et al. Template-assisted selective epitaxy of III–V nanoscale devices for co-planar heterogeneous integration with Si. Appl. Phys. Lett. 106, 233101 (2015).
2. Lee, S. et al. High Performance InGaAs Gate-All-Around Nanosheet FET on Si Using Template Assisted Selective Epitaxy. in 2018 IEEE International Electron Devices Meeting (IEDM) 39.5.1-39.5.4 (2018). doi:10.1109/IEDM.2018.8614684.
3. Borg, M. et al. High-Mobility GaSb Nanostructures Cointegrated with InAs on Si. ACS Nano 11, 2554–2560 (2017).
4. Mauthe, S. et al. InP-on-Si Optically Pumped Microdisk Lasers via Monolithic Growth and Wafer Bonding. IEEE J. Sel. Top. Quantum Electron. 25, 1–7 (2019).
5. Mauthe, S. et al. Hybrid III-V Silicon Photonic Crystal Cavity Emitting at Telecom Wavelengths. Nano Lett. (accepted).
6. Mauthe, S. et al. High-speed III-V nanowire photodetector monolithically integrated on Si. Nat. Commun. 11, 4565 (2020).

We report the successful growth of high-quality GaAs_1–xSb_x nanowires on monolayer graphene/SiO₂/p-Si (111) using molecular beam epitaxy (MBE) for the application of a flexible near-infrared photodetector. A systematic and detailed study of NW growth parameters, namely, growth temperature, V/III beam equivalent pressure (BEP) ratio, and Ga shutter opening duration, has been carried out. Growth of vertical <111> oriented nanowires on graphene with 4 K photoluminescence emission in the range 1.24–1.38 eV has been achieved. The presence of a weak D mode in Raman spectra of NWs grown on graphene suggests that NW growth did not alter the intrinsic properties of the monolayer graphene. High-resolution transmission electron microscopy and a selective area diffraction pattern confirmed the zinc-blende crystal structure of the NWs. This study suggests that Sb as a surfactant plays a critical role in the surface engineering of the substrate, leading to the superior optical quality of NWs exhibiting a higher 4 K photoluminescence intensity and lower full width at half maxima (fwhm) with significant improvement in optical responsivity compared to NWs grown on Si substrate of similar Sb composition. Novel growth conditions for improvement of NW density and device performance will be discussed.

InGaAs is an ideal material for near-infrared (IR) photodetection at telecommunications wavelengths, but its devices typically require the use of bulk III-V substrates which are costly and largely incompatible with CMOS processing.¹ Nanowire-based photodetectors are a promising route to the integration of III-V materials with Si electronics and even offer a number of performance advantages over bulk devices.² For example, the anisotropic geometry of nanowires allows for the possibility of polarization-resolved photodetection as well as fast carrier collection, making them suitable for fast and efficient multiplexed applications.³ Recently, arrays and networks of high-quality horizontal InAs and InGaAs nanowires have been demonstrated using selective area epitaxy on III-V substrates by growing them on top of GaAs nanomembrane buffers.^4,5 Extension of this technique through monolithic integration of GaAs nanomembranes on Si (001) offers the opportunity to create scalable, high-performance IR photodetectors at a fraction of the cost of current InGaAs-based devices. This, however, requires a reliable method to grow defect-free GaAs nanomembranes on Si. In this work, we investigate the growth of GaAs nanomembranes on Si by molecular beam epitaxy. Specifically, we combine various defect reduction techniques, including v-groove trenches, aspect ratio trapping and migration enhanced epitaxy, and characterize the resulting nanostructures. We demonstrate highly faceted nanomembranes several micrometers in length that exhibit large defect-free regions, which may be suitable as a template for future horizontal InGaAs nanowires. Furthermore, we introduce a novel architecture for photoconductive detectors based on these templated nanowires using plasmonic enhancement. Through frequency domain finite element simulations, we demonstrate polarization-sensitive absorption of up to 20% in the nanowire active material and the ability to tune the absorption response by changing the dimensions of the array. Finally, we present related work on the fabrication and characterization of photodetectors based on InAs_(1-x)Sb_x and InSb nanowire crosses and hashtags grown by the vapor-liquid-solid technique.^6,7,8 These structures offer the possibility of polarization-resolved photodetection in the mid-IR. To the best of our knowledge, IR photodetectors based on nanowire crosses and/or hashtags have not been previously reported.

References
1 R. LaPierre et al., J.Phys. D: Appl. Phys. 50 (2017)
2 C. L. Tan et al., Nanophotonics. 7 (2018)
3 A. Dorodnyy et al., IEEE J. Sel. Top. Quant. Electr. 24 (2018)
4 M. Friedl et al., Nano Lett. 18 (2018)
5 F. Krizek et al., Phys. Rev. Mat. 2 (2018)
6 S. A. Kahn et al., in preparation
7 D. Dalacu et al., Nano Lett. 13, (2013)
8 S. Gazibegovic et al., Nature. 548 (2017)
Acknowledgements: Funding from Swiss National Science Foundation, NCCR QSIT, EU H2020 funded INDEED network, Microsoft Station Q

In recent years, InAs quantum dot lasers have proven themselves resilient against crystalline defects that form during heteroepitaxial growth of III-Vs on silicon. The three-dimensional confinement and reduced in-plane diffusion lengths of carriers in quantum dots relative to quantum wells are the source of this resiliency, but alone these characteristics are not sufficient for demonstrating commercially reliable performance for lasers on silicon. Quantum dot active regions must be utilized together with highly optimized III-V/Si buffers to reduce the dislocation density to achieve acceptable device reliability.
In this talk, recent advances in GaAs/Si templates will be highlighted along with the resulting enhancements in laser reliability. Over the last three years, we have produced InAs quantum dot lasers on GaAs/Si templates with dislocation densities ranging from 3×10⁸ cm^-2 down to < 7×10⁶ cm^-2 by optimizing thermal cycling procedures and strained filter layers. As the material has improved, the extrapolated mean-time-to-failure has increased several orders of magnitude to >10,000,000 h at 35°C and >50,000 h at 60°C.
In addition to improving laser reliability, the improvements in material quality have virtually closed the performance gap relative to native substrate lasers before aging. At dislocation densities of 7×10⁶ cm^-2, low threshold, high power, and high efficiency lasers are readily achievable with continuous wave operation above 100°C. These improvements justified more advanced laser implementations and investigations into the dynamic performance of quantum dot lasers on silicon. The unique atom-like density of states of quantum dots has historically enabled unique properties such as an insensitivity to optical feedback, highly engineerable gain bandwidth, and ultrafast gain recovery; these properties unique to quantum dots have enabled Fabry-Perot lasers with complete insensitivity to optical feedback, high performance mode-locked lasers operating at 20 GHz with 64 channels capable of 4.1 Tbps transmission below the forward error correction (FEC) limit, and 100 GHz colliding pulse mode-locked lasers capable of 0.9 Tbps transmission below the FEC limit. Achieving each of these results required a high degree of optimization in the epi design and materials growth to obtain uniform distributions of high-quality quantum dots, which will be emphasized in the talk.
Further improving performance and, in particular, demonstrating commercially viable reliability at elevated temperatures >60°C will require additional efforts to reduce the dislocation density in the laser active region. Our efforts to push the threading dislocation density < 2×10⁶ cm^-2 while also reducing the buffer thickness will be discussed in addition to efforts to address misfit dislocations that have been observed in the laser active region. Given the much larger cross-section of material that interacts with a misfit dislocation relative to a thread, it is believed that addressing these defects is critical to reaching the goal of native substrate level device reliability on a silicon substrate.

Advancement in microelectronics technology enables autonomous edge computing platforms in the size of a dust mote, bringing efficient and low cost artificial intelligence computing close to end users. The key components of these edge computers include (i) communication devices to talk to nearby host computers and (ii) power sources to convert energy to on-chip electrical power. As RF communication and power converters suffer from low efficiency with reduced antenna size, optical devices can provide much higher efficiency in sub-centimeter footprint. As there is limited energy storage inside the small footprint of edge computers, we developed power-saving light-emitting-diodes (LEDs) with high efficiency at ultra-low current and voltage[1], by improving the quantum well design and increasing the radiative radiation rate at low injection current. The device exhibits close to unity internal quantum efficiency at a current density orders of magnitude lower than that for conventional LEDs. Wireless communication is demonstrated at these low power conditions. We also investigated dust-sized III-V photovoltaic (PV) cells grown on silicon and Silicon-on-insulator substrate[2], considering surface passivation and crystallographic defects. We demonstrated highest power density micro-PVs on silicon substrate. These silicon substrates with integrated III-V devices can be used as the substrate carrier for heterogeneous integration of various silicon chiplets using wafer-level-packaging.

N. Li et al. Nat. Photonics 13, 588–592
N. Li et al. Advanced Mat. to be published

Metastable crystal phases of conventional semiconductors comprise enormous potential for high-performance electro-optical devices, while at the same time benefitting from the chemical similarity to their stable counterparts which enables the reuse of established processing technology. AlAs, AlSb, AlP, GaP and Ge all possess indirect bandgaps in their thermodynamically stable cubic phase, whereas both theoretical calculations as well as first experimental reports suggest their band transition to be direct when their lattice periodicity is changed to the hexagonal wurtzite (WZ) or lonsdaleite (LD) symmetry [1, 2]. These findings recently triggered intense research interest for creating efficient light sources in the important amber-green wavelength regime of the visible spectrum where a lack of suitable emitting materials currently limits the performance of LEDs and semiconductor lasers [3]. Beyond that, direct bandgap Ge and SiGe compounds could even pave the way towards active group-IV optoelectronic devices. However, synthesizing these novel polytypes remains challenging and so far, has mainly been achieved in the form of thin nanowires, mostly by using Au catalysts and by compromising other important parameters such as crystal morphology or doping, hampering scientific and commercial exploitation.

In this work, we demonstrate new techniques to synthesize layers in their metastable WZ phase. InP is particularly suited for this task as both the stable cubic and the WZ phase exhibit a direct, but distinctive bandgap, which allows efficient optical analysis. We use MOCVD and selective area epitaxy to grow pure WZ nanowires and fins. We then show two extensions of this technique to obtain planar layers: The first one is based on confined epitaxy and enables growth on standard (001)-oriented substrates [4]. In a second, newly introduced approach we explore epitaxial layer overgrowth (ELO) on (111)A-oriented wafers. This allows to grow pure WZ layers on insulator exceeding areas of 100 μm², constituting a promising substrate for device fabrication. The material quality of the structures is determined by micro-photoluminescence (μ-PL), high-resolution scanning transmission electron microscopy (HR-STEM), atomic force microscopy (AFM), and cathodoluminescence (CL). We compare the investigated techniques, show their limitations and develop a general model to explain polymorphism in planar layers.

This work was supported by the EU H2020 program SiLAS (Grant Agreement No. 735008).

[1] De, A. et al. Phys. Rev. B 81, 155210 (2010).
[2] Cartoixà, X. et al. Nano Lett. 17, 4753–4758 (2017).
[3] Gagliano, L. et al. Nano Lett. 18, 3543–3549 (2018).
[4] Staudinger, P. et al. Nano Lett. 18, 7856–7862 (2018).

One of very important requirements for nanolasers intended for future photonic chips is the energy consumption. The amount of total energy input to a laser required to a large extend is determined by the critical value of carrier density, at which optical gain occurs. In conventional semiconductors, optical gain is based on electron-hole plasma and it occurs at the so-called transparency density, which is near and above the Mott density, typically on the order of 10¹² per square centimeter. In 2D materials, the orders of magnitude larger exciton binding energy make it possible to have optical gain based on various excitonic species such as excitons, trions, bi-excitons etc. Such optical gain can occur at density levels several orders of magnitude smaller, well below the Mott density. In this talk, we will discuss possibilities of such optical gain in 2D materials that could occur at extremely low injection levels. As an example, we will discuss our recent experimental results [1] on trion gain in a gated 2D MoTe2 structure that occurs at the density level ~5 orders of magnitude smaller than the Mott density.

References:
[1] Z. Wang, H. Sun, Q. Zhang, J. Feng, J. Zhang, Y. Li, C.Z. Ning, Excitonic Complexes and Optical Gain in Two-Dimensional Molybdenum Ditelluride Well below Mott Transition, arXiv:1812.04296 [cond-mat.mes-hall]

Photonic integrated circuits offer high-speed data transmission and reduced energy per bit consumption. To meet the energy requirements for on-chip communication, a small form factor of the integrated III-V semiconductor emitters and detectors is crucial. Ding et al. suggest a laser volume of less than ~2(λ/n)³ to lower the energy to 10 fJ/bit.¹ Scaling to these dimensions, however, is not possible for purely photonic devices, hence hybrid plasmonic-photonic cavities are being explored.²
Using template-assisted selective epitaxy (TASE), we can integrate high quality III-Vs epitaxially on Si, despite their lattice and thermal mismatch.³ We have demonstrated InP microdisk lasers grown via TASE⁴ and are currently integrating InGaAs quantum wells⁵ for emission above the Si absorption edge. Moreover, we have used intensity measurements of scattered photoluminescence (PL) in Au waveguides on InP to confirm that PL couples to surface plasmons and is guided in this materials system.⁶

In this work, we explore Au cladded III-V whispering gallery mode (WGM) cavities to further scale down the footprint of lasers. We design and fabricate InP cavities with resonances in the near-infrared using direct wafer bonding. InP is bonded on Si and (WGM) cavities with diameters from 100nm to 1000nm are dry etched subsequently. In a second step selected cavities are cladded with Au and the performance of purely photonic and metal-clad devices is compared. μ-PL spectroscopy is performed using a ps-pulsed supercontinuum laser at 750nm. A 100× objective is used to focus light on the sample from the top and collect the PL response. Measurements show evidence of lasing with resonant emission around 900 nm and a far field radiation pattern with fringes for cavities above 500 nm in diameter in both cases, with and without Au. For the Au cladded case, the threshold is higher. Indication of lasing is not observed for 400 nm wide purely photonic structures. However, for the cavities cladded with Au, a resonant peak emerges, indicating lasing at around 920 nm. Finite difference time domain simulations fit with our experimental observations and confirm a decrease in quality factor with decreasing diameter for the photonic cavities. Around 600 nm diameter, Q-factors become comparable to Au cladded cavities. When further reducing the cavity size, purely photonic cavities support no resonances anymore while cladded cavities do.

In summary, we observe evidence of lasing at room temperature for WGM cavities upon pulsed optical excitation. Preliminary results show that for smaller cavities one benefits from metal cladded designs and is able to scale down the emitter beyond the limit of purely photonic cavities. In a next step we will apply the same approach to InP cavities directly grown in silicon using TASE. This might further improve performance as the TASE growth results in atomically smooth sidewalls with reduced defect density.⁴

This work was supported by H2020 ERC project PLASMIC #678567 and by the Korea MSIT #2017R1A2B4007219.

1. Ding, K. et al., Las&Phot Rev, 9, No. 5, 488 - 497 (2015).
2. Oulton, R. et al., Nat. Phot. 2, 496–500 (2008).
3. Schmid, H. et al., Appl. Phys. Lett. 106, 233101 (2015).
4. Mauthe, S. et al., IEEE JSTQE. 25, 1–7 (2019).
5. Baumgartner, Y. et al., IEEE 14th International Conference on Group IV Photonics, Berlin (2017).
6. Tiwari, P. et al., 10^th international conference on Metamaterials, Photonic Crystals and Plasmonics, Lisbon (2019).

With ever-increasing demand for faster speeds, copper interconnects that ferry vast amounts of data in computing devices are expected to significantly lag behind the requirements, thereby creating a bottleneck for moving data rapidly. Optical interconnects provide a promising alternative, allowing much faster speeds and larger bandwidths between different components, on or off chip. Three most critical interconnect components are light sources, waveguides and detectors. Currently, optical communication is performed by a combination of these devices where the information is encoded in photon number and frequency (intensity and bandwidth). However, no information is encoded in intrinsic properties of photons, i.e., photon spin and also possibly the orbital angular momentum modes of the light beam. Analogous to the idea of spintronics, where the spin degree of freedom of electrons are utilized to encode and transmit information, similar enhancements in the data processing capabilities of the optical interconnect systems can be obtained if information can be encoded in photon spin. Photons, have two fundamental spin polarizations reflected in their polarization state, i.e., right and left circular polarization. To fully utilize more degrees of freedom of light to enhance the data processing capabilities of network systems, it is required that new light sources, detectors and waveguides be designed and fabricated that can produce, transmit and detect light with complex polarizations. We will discuss some recent developments in our laboratory towards the development of topological waveguides that can route photonic signals based on photon spin and also photodetectors that are sensitive to circular polarization (photon spin). Either by protecting or breaking certain symmetries in (meta)materials, new photonic materials and devices will be discussed that can enable a new generation of photonic devices that can encode, transmit and sense information encoded in spin and angular momentum modes of light that are promising for the development of integrated chiral photonic systems with extremely large data processing capabilities.

Artificial Intelligence (AI) is transforming our lives in the same way as the advent of the Internet and cellular phones has done. AI is revolutionizing the healthcare industry with complex medical data analysis, actualizing self-driving cars, and beating humans at strategy games such as Go. However, it takes thousands of CPUs and GPUs, and many weeks to train the neural networks in AI hardware. Traditional CPUs, GPUs, and neuromorphic (i.e., brain-inspired) electronics such as the IBM TrueNorth and Google TPU, may find it a challenge to train the neural networks of the future.

Neuromorphic photonics has experienced a recent surge of interest over the last few years, promising orders of magnitude improvements in both speed and energy efficiency over digital electronics using: artificial neural networks, spiking neural networks, and reservoir computing. By combining the high bandwidth and efficiency of photonic devices with the adaptive, parallelism and complexity attained by methods similar to those seen in the brain, photonic processors have the potential to be at least ten thousand times faster than state-of-the-art electronic processors while consuming less energy/computation.

We will provide an overview of neuromorphic photonic systems and their application to optimization and machine learning problems. We will discuss the physical advantages of photonic processing systems and describe underlying device models that allow practical systems to be constructed. We also describe several real-world applications for control and deep learning inference. Lastly, we will discuss scalability in the context of designing a full-scale neuromorphic photonic processing system, considering aspects such as signal integrity, noise, and hardware fabrication platforms. This talk is intended for a wide audience and hopes to teach how theory, research, and device concepts from neuromorphic photonics could be applied in practical machine learning systems.

Here we propose and emulate a photonic neural network whose neuron’s non-volatile weighting functionality is realized through an engineered hybrid Ge₂Sb₂Se₄Te₁-silicon Mach Zehnder modulator photonic memory with thermoelectrical programmability. The network can effortlessly perform inference tasks with high accuracy at the speed-of-light.

Machine-learning tasks performed by neural networks demonstrated useful capabilities for producing reliable, and repeatable intelligent decisions. Integrated photonics, leveraging both component miniaturization and the wave-nature of the signals, can potentially outperform electronics architectures when performing inference tasks. However, the missing photon-photon force challenges non-volatile photonic device-functionality required for efficient neural networks. Here we present a novel concept and its optimization of multi-level non-volatile photonic memories based on an ultra-compact (<4µm) hybrid phase-change-material Ge₂Sb₂Se₄Te₁-silicon Mach Zehnder modulator, with low insertion losses (3dB), to serve as node in a photonic neural network (NN). An optimized electro-thermal switching mechanism is engineered, causing phase transitions in the GSST, induced by heating by tungsten contacts, which allows weight updating functionality of the network. We show that a 5 V pulse-train (<1µs, 20 pulses) applied to a serpentine contact produces crystallization and a single pulse of longer duration (2µs) amorphization, used to set the analog synaptic weights of a neuron. Emulating an opportunely trained 100 x 100 fully connected multilayered perceptron NN with these weighting functionalities embedded as photonic memory, shows a 92% inference accuracy and robustness towards noise when performing predictions of unseen data. Exploiting the low delay interconnectivity of photonic integrated chips and the non-volatile transitions of PCM, an all-optical (AO) trained NN, that effortlessly performs dot-product functionality can be achieved, enabling intelligent computing functionality at the time-of-flight of the photon.


The written portion of the material corresponds to discrete states of the weighting function, and assuming a stabile writing resolution of about 500 nm, same achieved by optical writing, the total amount of available quantized resolution is given by 8 distinct states (3-bit), which can be further improved by extending the device length by multiple of L_π. Additionally, this is a reversible process, which allows to update the weights after many execution times. Interestingly, this solution is not hindered by insertion losses, which are negligible due to the rather low absorption coefficient of GSST at 1550 nm, and the total losses (~3dB) are mainly caused by the balancing mechanism (in this first analysis straightforwardly obtained achieved by placing a gold contact on the balancing arm). As an interim conclusion, this novel PCM MZM features by a micrometer-compact (<4µm) footprint and low insertion losses (3dB), enabling the implementation of a deep NN which comprises multiple nodes.

As preliminary study, we estimated the functionality of the proposed perceptron as main unit of the NN, by emulating its behavior in a 3-layer fully connected NN implemented in the Google Tensorflow tool and, as an initial example, for the MNIST data set. Nonlinear activation functions (here considered as electro-optic) are placed between two consecutive layers on each input connection. The network is trained both without and with noise of the weights and NLAF. Our hypothesis, confirmed by preliminary studies on the network is that, when we allow for a certain amount of noise during the training, the model during the inference stage, becomes more robust and the network can perform inference with up to 92% accuracy showing robustness towards noise.

Realization of on-chip scalable optical quantum information processing (QIP) systems requires optical circuits built around arrays of single photon sources (SPSs) to manipulate the emitted photons and enable interference between photons from distinct SPSs resulting in entanglement. Towards such a goal we have reported a new class of on-chip single quantum dot arrays - grown on lithographically fabricated mesas designed to exploit the attendant surface stress gradient driven adatom migration during MBE growth thus termed mesa-top single quantum dots (MTSQDs) [1]. These MTSQDs, we have demonstrated, are ~99% pure SPSs [2] that are readily integrable with on-chip light manipulating units (LMUs) that provide the needed multiple functions of enhancing the emission rate of the SPS and the directionality of the photon emission in the horizontal direction, on-chip propagation, splitting and recombining to enable on-chip interference towards entanglement.
So far attempts towards creating such LMUs (mostly resonant cavity and waveguide) have been based on photonic crystal membrane structures and exploit departure from Bragg scattering to provide localized photon modes for individual cavity or waveguides thus facing the challenge of strict mode-matching between these components. In contrast, in this talk we present an approach to realizing all the required light manipulating functions based on metastructures made of subwavelength size dielectric building blocks (DBBs) where a common collective Mie resonance of the interacting DBBs is exploited to provide all the needed light manipulating functions simultaneously [3, 4, 5]. These Mie resonances are fundamentally different from Bragg scattering as they are typically broad in spectrum, with a typical Q~100- that alleviates the strict requirements of spectral and spatial matching between the network components and with the SPS emission mode. Moreover, Mie resonances are less sensitive to effect of fabrication disorders that introduce aperiodicity in the array. Furthermore, the spectral broad nature of the Mie resonances allows controlled interference between the electric and magnetic modes resulting in directionality of photon emission and propagation without strong field localization—an effect that is unique to this class of LMUs. Importantly, providing all the needed light manipulating functions using the same collective Mie mode circumvents issue of mode mismatch between the components of the optical circuit including the SPS, and thus allows design and realization of large-scale circuits.
Aimed at the above noted objective of on-chip nanophotonic systems, in this talk we will present finite element method (FEM)-based design and simulation of the response of an MTSQD-DBB integrated optical circuit that exploits the dominant collective magnetic and electric dipole mode to provide the MTSQD emission rate enhancement (Purcell effect) ~5 as well as emission directionality (the nanoantenna effect), lossless on-chip propagation, beam-splitting, and beam-combining [4, 5]. We show that such structures enable on-chip interference of photons resulting in path entanglement between the two MTSQD SPSs at large distances.
Finally, we will present FEM simulations of direct coupling of two SPSs via an intermediate lossless collective Mie mode over on-chip distances much longer than the wavelength. We show that such coupling can result in a super-radiant state involving on-chip SPSs at a distance resulting in a ~2 fold enhanced decay rate [5]. This emergent super-radiant state in such SPS-SPS coupled system is maximally entangled, and thus may act as a potential resource for on chip QIP.

[1] J. Zhang et.al, J.Appl.Phys.120,243103(2016)
[2] J.Zhang et. al, App. Phys. Lett. 114, 071102(2019)
[3] S. Chattaraj, J. Opt. Soc. Am. B. 33, 12(2016)
[4] S. Chattaraj et.al, arXiv1811.06652v1(2018)
[5] S. Chattaraj et.al, IEEE J. Quant. Electron. 2019, Accepted for Publication
This work is funded by ARO W911NF-15-1-0025

Integrated photonic circuits efficiently combine multiple optical functions on a single physical platform, such as guiding, modulating, splitting, or detection of light. The integration of these functions removes barriers for designing and realizing large optical circuits, as present when using discrete components like bulky lenses or mirrors in tabletop optical networks. With the rise of silicon photonics, integrated photonic circuits (PICs) are becoming increasingly large and highly functional, eventually allowing unprecedented concepts of photonic computing. Examples are integrated photonic quantum processors, microwave photonic filters and processors, and optical accelerators to train and execute neural networks.

An important building block in integrated optical circuits is an efficient link between the optical and electrical domain. Well-known examples of such links are integrated high-speed modulators to convert electrical signals into optical signals at very high-speed, and low-power tuning elements to compensate for variations in the device operation temperature and for device-to-device variations during fabrication. To enable such electro-optic links, the two most widely used physical effects are the plasma-dispersion effect and Joule heating. Although these effects are attractive to use due to their compatibility with standard photonic fabrication processes, their performance in integrated devices is intrinsically limited by high insertion losses and high-power dissipation.

Over the past decade, we established an alternative electro-optic switching technology by embedding a Pockels material into silicon-based photonic devices [1]. We reached this goal by developing a process to fabricate ferroelectric barium-titanate (BTO) thin films on silicon substrates using advanced epitaxial deposition techniques and by developing a BTO process technology [2]. We correlated the electro-optical properties of the thin films with their structural properties such as porosity and crystalline symmetry to show guidelines for improving the functional properties. By realizing integrated hybrid BTO/silicon devices, we demonstrated record-high, in-device Pockels coefficients of >900 pm/V. The Pockels effect in BTO-based photonic devices indeed enables extremely fast data modulation at rates beyond >40 Gbps and ultra-low-power electro-optic tuning of silicon and silicon-nitride waveguides. We also show ways of how to integrate and use BTO in plasmonic slot waveguide structures for very compact optical devices. With the development of a wafer-level integration scheme of single-crystalline BTO layers to a 200 mm process [3], we could demonstrate a viable path to combine the BTO-technology with existing fabrication routes.

With major breakthroughs in the past years, BTO has emerged as a strong candidate for a novel generation of electro-optic devices. Major achievements of the BTO technology will be covered in the presentation, ranging from important materials aspects, device development, integration concepts, and novel applications in the area of quantum computing, high-speed communication, and neuromorphic optical computing.

[1] S. Abel et al., “Large Pockels effect in micro- and nano- structured barium titanate integrated on silicon,” Nat. Mater., vol. 18, no. 1, pp. 42–47, 2019.
[2] S. Abel et al., “A strong electro-optically active lead-free ferroelectric integrated on silicon,” Nat. Commun., vol. 4, p. 1671, 2013.
[3] F. Eltes et al., “A BaTiO3-Based Electro-Optic Pockels Modulator Monolithically Integrated on an Advanced Silicon Photonics Platform,” J. Light. Technol., vol. 37, no. 5, pp. 1456–1462, 2019.

In this work, we find precision layered-material optical designs that can perform state-of-the-art edge detection among compact-form-factor approaches, offering high efficiency and large angular bandwidth / numerical aperture. We identify these designs by two approaches: first, adjoint-based “inverse design” of the layered structures, for rapid optimization of many parameters, and second, by training the structures on software-based edge-detection algorithms that can be “learned” into the hardware itself. Optical computing platforms based on such layered media offer simpler integration and more compact form factors than structured or free-space alternatives. We identify the benefits and tradeoffs of ideal material properties—metal/dielectric, lossy/lossless—within these optimal designs, and discuss its generalization beyond edge detection to more general optical analog computing.

Insects are capable of amazing tasks, such as traveling across hundreds of kilometers of unfamiliar terrain to pinpoint a specific breeding ground, with a brain that contains 100 000x fewer neurons than our own brains, using merely a few drops of nectar as energy supply. At each moment during navigation, the insect has to decide whether to turn right, left or go straight. Integrating sensory inputs, internal states, and previous experience, it computes its current and intended headings, compares the two, and initiates steering. A brain region called the central complex (CX) is essential for these tasks. Recently a biologically constrained computational model of the CX has been generated that can carry out these computations to generate homing behavior [1]. This computational model has been shown to carry out homing tasks using speed and heading information of limited precision, with considerable noise in the circuit. Containing less than 100 neurons, it can serve as a model system to understand the functionality and prove the efficiency of novel hardware solutions.

To realize this CX model we propose and theoretically model a network circuit based on existing III-V nanowire (NW) opto-electronic components that are optically connected through a shared waveguide structure. The neural weights of the mutual interconnections are achieved by controlling the far-field light emission patterns and by strategic placement of the nano-components. As a result, there is no need for physical connecting wires between the network nodes. Each node in this network is capable of both receiving, evaluating and sending optical signals, while power consumption is minimized by using high efficiency concepts from NW based photovoltaics and quantum dot emitters [2-6]. We show that the modeled coupling efficiencies and activation functions well suffice to facilitate excellent navigation in the insect neuron model [1]. As the efficiencies and electrical performance of the NW devices are known, we can estimate the energy consumption per operation and find that this should be at least as low as in biological processes [7] and beyond standard hardware implementations.

Our detailed simulations of the fundamental device component, which will act as a basic sigmoid neural node throughout the network, show how it performs the tasks of receiving light inputs, evaluating them and emitting a light output. We model it as an electrical circuit which is parametrized and solved by a standard SPICE solver (Ngspice). We describe the subcomponents using an Ebers-Moll model and the Shockley diode equation, respectively, where the relevant properties were calculated using a drift-diffusion model with thermionic boundary conditions in COMSOL. A NW phototransistor was optimized to give enough gain to broadcast the signal to the next layer using a realistic non-radiative Shockley-Read-Hall recombination lifetime for NWs of 1.34 ns [8] as well as Lorenz-Mie scattering theory. Extensive finite difference time domain (FDTD) simulations (Lumerical), were made for the central part of the CX model which consists of eight fully interconnected components. They were placed in a shared asymmetric waveguide which was optimized for the target wavelength. As the neurons of this layer are interconnected, input-output isolation is non-trivial, and the FDTD model was used to optimize the network in this respect. The gain of the fundamental device ensures both cascadeability and proper fan-out without the need of critical biasing [9], and the analog architecture carries a built-in robustness to noise [1].
[1] T. Stone, et al, Curr. Biol. 27, 3069 (2017)
[2] J. Claudon, et al, Nat. Photonics 4, 174 (2010)
[3] M. E. Reimer, et al, Nat. Commun. 3, 737 (2012)
[4] J. Wallentin, et al, Science 339, 1057 (2013)
[6] G. Otnes and M. T. Borgström, Nano Today 12, 31 (2017).
[7] C. Mead, Proc. IEEE 78, 1629 (1990)
[8] H. J. Joyce, et al, Nano Lett. 12, 5325 (2012)
[9] D. A. B. Miller, Nat. Photonics 4, 3 (2010)

Nanophotonic networks, where light flows in a mesh of interconnected optical waveguides, can serve as an unconventional lasing cavity, supporting resonant modes that are formed by light propagation and interference across many network links. In this talk, I will discuss our recent efforts to make spectrally controllable lasers with nanostructured sub-wavelength networks [1]. The networks are made from polymer nanofibers that are doped with a laser dye and can lase over a wide range of frequencies, corresponding to many network modes. Interestingly, we can exploit mode competition in the network laser for spectral control and signal processing. For example, we can experimentally single-out a lasing frequency out of the spectral haystack by selectively illuminating a very small subset of the network links. This result is also predicted by a graph description of Maxwell’s equations, and a steady-state ab initio laser theory (netSALT) which is used to model the network laser. In particular, netSALT enables the design of the network topology for lasing and identification of the most important edges in the network that are necessary for sustaining a particular mode.

[1] A nanophotonic laser on a graph, M.Gaio, D. Saxena, J. Bertolotti, D. Pisignano, A. Camposeo, R. Sapienza, Nature Comm. 10, 226 (2019)

We introduce a physical mechanism to perform machine learning by demonstrating a Diffractive Deep Neural Network (D²NN) architecture that can all-optically implement various functions following the deep learning-based design of passive layers that work collectively. We created 3D-printed diffractive networks that implement classification of images of handwritten digits and fashion products as well as the function of an imaging lens, spectral filters and wavelength demultiplexers at terahertz spectrum. This passive diffractive network can perform, at the speed of light, various complex functions that computer-based neural networks can implement, and will find applications in all-optical image analysis, feature detection and object classification, also enabling new camera designs and optical components that perform unique tasks using diffractive neural networks designed by deep learning.

The success of deep learning has recently stimulated broad interest in optical neural networks due to their key advantages: (1) ease of parallelization and scalability, (2) low power and (3) high data processing speed. Merging wave optics with deep learning methods, Diffractive Deep Neural Networks (D²NN) have been introduced as an optical neural network framework, which is based on the engineering of light-matter interaction occurring at successive diffractive layers that connect an input (e.g., an object to be classified) to an output plane through optical diffraction. During the design of a D²NN, which is performed in a computer using deep learning methods, including e.g., error backpropagation and stochastic gradient descent, the phase and/or amplitude patterns of each diffractive layer of a network are jointly-optimized to solve a given computational task such as image classification. After its design, the resulting diffractive optical network is physically fabricated to perform all-optical inference at its output plane. Successful demonstrations of diffractive neural networks include all-optical object classification and ultra-compact imaging systems, involving 3D-printed diffractive layers that compute through the diffraction of terahertz (THz) radiation. In these previous works, the diffractive networks processed spatially and temporally coherent radiation, involving a monochromatic plane-wave illumination. Here we expand our D²NN framework to simultaneously process a continuum of wavelengths, making it applicable to all-optical computing settings where broadband illumination is essential and/or advantageous. Using the native and/or engineered dispersion of materials, broadband diffractive optical neural networks simultaneously process all the target wavelengths within a desired band to compute a statistical or deterministic task that it is trained for. We experimentally confirmed the success of this framework by designing, 3D-printing and testing seven different diffractive networks that process an input THz pulse to perform a series of tunable, single passband and dual passband spectral filters, as well as spatially-controlled wavelength de-multiplexing. Using a THz Time-Domain Spectroscopy (TDS) system, we demonstrated a very good match between our experimental results and the corresponding numerical design for each fabricated broadband diffractive network. Leveraging deep learning-based optimization tools and material dispersion properties, this broadband diffractive network framework provides a powerful method to engineer light-matter interaction in 3D, opening up new avenues to design non-intuitive, task-specific optical components as well as all-optical classifiers and processors. The presented method is broadly applicable to various parts of the electromagnetic spectrum, including the visible band, and can be further strengthened by the additional degrees of freedom provided by engineered material systems including e.g., metamaterials. We believe broadband diffractive networks provide a unique framework to perform all-optical statistical inference for machine learning tasks as well as to design deterministic and task-specific optical components, broadly covering a non-intuitive and yet powerful design space.

Optical analog computing provide solutions with fast and power-efficient performance for big-data applications. Recently investigated metamaterials and metasurfaces demonstrate potentials in mathematical operations over wavelength-scale footprint. We extend the functionality of using nonlocal dielectric metasurface to perform spatial differentiation equation for imaging processing. The study pave the way towards ultrathin devices for optical signal processing.

Image processing has become a critical technology in a variety of science and engineering disciplines. While most image processing is performed digitally, optical analog processing has the advantages of being low-power and high-speed though it requires a large volume. Here, we demonstrate optical analog imaging processing using a flat optic for direct image differentiation allowing one to significantly shrink the required optical system size. We first demonstrate how the image differentiator can be combined with traditional imaging systems such as a commercial optical microscope and camera sensor for edge detection at a numerical aperture up to 0.32. Second, we demonstrate how the entire processing system can be realized as a monolithic compound flat optic by integrating the differentiator with a metalens. The compound nanophotonic system manifests the advantage of thin form factor optics as well as the ability to implement complex transfer functions and could open new opportunities in applications such as biological imaging and machine vision.

The differentiator consists of carefully designed silicon nanorods that can transform an image into its second-order derivative, taking the Laplacian of the incident field. The optical response is based on interference between direct transmission and a low-Q guided resonance allowing for transmission at large incident angles but none at normal incidence. Fourier imaging was carried out showing th required quadratic transfer function for an NA up to 0.32, which allows resolving edges at the scale of 4 microns. To showcase its practical application, we first demonstrate the device at a wavelength of 740 nm by integrating the differentiator with a commercial microscope. This allowed us to reveal high-contrast edge information when imaging biological cells, akin to phase contrast microscopy though with a considerably simpler system. Using scalable manufacturing techniques, not involving lithography, we also fabricated the nanoarray over a centimeter-scale wafer and demonstrate edge detection when placing the filter directly in front of an imaging sensor. Lastly, we have demonstrated the creation of a monolithic flat optic combining both a metalens and the differentiator for realizing a compact image processing system.

Combinatorial optimization problems are ubiquitous in our modern life. Classical examples include lead optimization in drug and biocatalyst discovery, resource optimization in wireless communications, logistics and scheduling, sparse coding for compressed sensing, deep machine learning in artificial intelligence and fintech. These optimization problems can be mapped to either Ising model, XY model or k-SAT problem, which is a main reason why various Ising, XY and SAT solvers have been proposed and implemented in the recent years.

We have focused on the network of optical parametric oscillators to construct coherent Ising machines, XY machines and SAT solvers. An optical parametric oscillator (OPO) operates as a quantum analog device at below oscillation threshold and also functions as a classical digital device at above oscillation threshold. We need not only quantum computational resources, such as quantum correlation and quantum suppression of chaos, but also classical computational resources, such as spontaneous symmetry breaking and exponential amplitude amplification, to build an efficient optimizer. An OPO is almost a unique choice of device to realize such quantum and classical computational resources simultaneously at room temperatures.

A study on quantum state engineering in an OPO with measurement-feedback control was dated back to 1980s [1,2]. A PPLN waveguide OPO device was identified as a stable and efficient generator for squeezed vacuum state pulses at communication wavelengths in 1990s [3]. An idea of using the pitchfork bifurcation and spontaneous symmetry breaking at OPO threshold as an irreversible decision making process was proposed in 2013 [4] and experimentally implemented [5,6]. A scalable architecture based on measurement-feedback coupling scheme was subsequently demonstrated [7,8]. Those devices already demonstrated competitive performance against modern heuristics [9] and quantum annealers [10]. In this talk, we will discuss the principles of quantum-classical crossover and future prospects of quantum neural network.

[1] G. Bjork and Y. Yamamoto, Phys. Rev. A 37, 4229-4239 (June 1988).
[2] K. Watanabe and Y. Yamamoto, Phys. Rev. A 38, 3556-3565 (October 1988).
[3] D. K. Serkland et al., Opt. Lett. 20, 1649-1651 (August 1995).
[4] Z. Wang et al., Phys. Rev. A 88, 063853 (December 2013).
[5] A. Marandi et al., Nature Photonics 8, 937-942 (October 2014).
[6] T. Inagaki et al., Nature Photonics 10, 415-419 (June 2016).
[7] P. L. McMahon et al., Science 354, 614-617 (October 2016).
[8] T. Inagaki et al., Science 354, 603-606 (October 2016).
[9] T. Leleu et al., Phys. Rev. Lett. 122, 040607 (February 2019).
[10] R. Hamerly, Science Advances 5, eaau0823 (May 2019).

Physical unclonable functions (PUF) are considered as one of the most promising methods for hardware authentication. PUFs are typically compact devices providing a response when challenged that is linked to uncontrollable stochastic processes and physics variability that occur during hardware manufacturing. As a result, it is practically impossible to produce two identical devices, which is highly desirable for applications in cybersecurity. For example, an important source of statistical variability in semiconductor nano-electronics, which has a huge impact on the electrical characteristics of each device, is the random dopant distribution (RDD) within the structure. The stochastic nature of the RDD profile for each device leads to a unique electrical output for each device. Indeed, this is the main idea and strength behind PUFs based on quantum dots that can be fabricated using standard CMOS technology and processes. In addition, as the dimensions of the modern transistors range from a few to tens of nanometers, there is a unique opportunity to use quantum mechanical effects, such as confinement and tunneling, as another layer of security based on quantum effects. One possible device structure that can be fabricated using the standard CMOS process and that can be implemented in current electrical circuits is a quantum-dot-in-wire device, which is considered in this work. The dot device relies on tunneling of electrons through a barrier and for this reason it can be considered a device that provides a quantum fingerprint.

Our work shows quantum mechanical simulations of quantum-dots (QDs) embedded within Si and III-V nanowires. To capture the effect of statistical sources of variability, we simulated more than 100 automictically unique nanowire devices with differing numbers and positions of dopants, not only in the quantum dot but also at the source and the drain regions also. Our results reveal that the specific number of dopants and their positions give rise to unique current-voltage characteristics, providing unique signatures for use as the basis of physical unclonable functions (PUFs). Hence, unique current-voltage characteristics characteristics can be defined as quantum fingerprints and applied to problems in cybersecurity, such as authentication and identification. Here we have also established a link between the RDD position, the transmission spectra, current spectrum and local density of states. Our work captures the complex nature of the quantum effects in such ultra-small-scaled devices and it can be used for investigating quantum mechanical effects in not only the dot but also in conventional transistors. Moreover, adoption of hardware security devices for authentication is on the rise. The technology proposed here delivers a practical means to extract fingerprints from quantum confined systems that could provide robust security to current electronics.

In recent years, Ising model solvers based on artificial spin systems have been studied intensively. A coherent Ising machine (CIM) is one of such systems in which the binary phase state of degenerate optical parametric oscillators (DOPO) is used to represent an Ising spin. By using long-distance (typically 1 km) fiber cavity that contains a phase sensitive amplifier based on the 2^nd or 3^rd order nonlinear optical effect, we can generate thousands of time-multiplexed DOPO pulses. The “spin-spin” interaction is implemented by using a measurement-feedback scheme, with which we can effectively realize all-to-all connection among thousands of DOPO pulses. The networked DOPO pulses are most likely to oscillate at a phase configuration that best stabilize the whole network, which gives the solution to the given Ising problems. In this talk, I will describe the large-scale CIM developed by NTT, which could find solution to maximum cut problem of 2000-node graphs faster than a conventional digital computer.

Light-matter coupling can e.g. influence chemical reactivity of molecules or modify transfer of energy through matter via formation of hybrid light-matter polaritonic states [1]. To elucidate the microscopic nature of such effects, we extend quantum electrodynamical density functional theory (QEDFT)--a first-principles, non-perturbative methodology for describing interactions between quantized light and matter [2]--by fully accounting for photonic losses in realistic optical cavities. We demonstrate that our method successfully addresses situations where the photonic losses dominate the dynamics of the excitations and show that the first-principles approach correctly predicts spectral response of such weakly coupled systems. We further apply the method to describe the cavity-induced interaction among electronic excited states, opening up new pathways in strongly-coupled molecular systems.

This work was supported by the DOE ‘Photonics at Thermodynamic Limits’ Energy Frontier Research Center under grant DE-SC0019140. DW is an NSF Graduate Research Fellow.

[1] Flick, J., Rivera, N., & Narang, P., Nanophotonics, 7 (9), 1479-1501 (2018).
[2] Flick, J. & Narang, P., arXiv:1907.04646v1 (2019).

A Photonic CMOS Field Effect Transistor includes an ultra low-resistance, high quantum efficiency laser fabricated in the MOSFET drain region, and multiple photon sensors or Avalanche Photo Diodes (APD) in the channel / well regions. The MOSFET, laser, and APD's are fabricated as one integral transistor. When a voltage is applied to the gate and a voltage is applied to the drain, both MOSFET and laser are turned on. Light emitted from the laser is absorbed by the APD to cause an avalanche breakdown. The high breakdown current flows into the drain to be a high output current. When the MOSFET is off, both laser and APD are switched off. Very low resistance and forward voltage lasers (VF near 0V) have been reported. With modulation doped junctions, bandgap engineered tunneling mechanisms, and low temperature selective epitaxy process (III-V or II-VI direct bandgap semiconductors on Silicon), these very low resistance lasers or LEDs are suitable for being fabricated in the drain of the Photonic CMOSFETs.

The Photonic CMOSFETs not only generates an internal laser light, but also can be used as an external light sensor. The internal and external lights may be of the same or various different frequencies. In the former case, it is defined as Chromatic Photonic CMOS. In the latter case, it is defined as Panchromatic Photonic CMOS. The photon sensor, or avalanche photo diode in the Chromatic Photonic CMOSFETs is designed to absorb lights from a specified range of light spectrum. Narrow bandgap semiconductors, or wide bandgap semiconductors can be used to build photon sensors in the Photonic CMOSFETs for photon absorption from infrared to ultraviolet ranges.

Three light signals are processed by a Photonic MOSFET. The first signal comes from the drain voltage. The second signal comes from the gate voltage. The third signal comes from the external light. Light emitted from the built-in laser in the drain region is modulated by the gate and drain voltages, and absorbed by a specific photon sensor in the channel / well region. The external light is absorbed by another photon sensor in the well region, and generates a different light current. All these three light signals are processed and multiplexed by the Photonic CMOS transistor.

Nonlinear optical films can be fabricated in the well region of a Photonic MOSFET.
The frequency or color of an input light signal is different from the output light due to multiple photon generations and absorptions in nonlinear optical films. The purpose of the nonlinear optical films is to increase the photon absorption rate and internal quantum efficiency of a Photonic MOSFET, and improve the laser multiplexing and communications. The nonlinear optical crystals can be fabricated on top of a total reflection layer (metal, or multiple refraction indices films), and below the CMOS channel region.

We will present information for device configurations of the Chromatic and Panchromatic Photonic CMOSFETs, and illustrate how light signals are modulated and multiplexed. We also include nonlinear optical crystals - and show the effects of such various nonlinear optics on the signal processing and light generation mechanisms. We will also provide evidence how the Chromatic Photonic CMOS technology may offer additional benefits for laser communication and more efficient signal processing systems.

Information regarding the Photonic CMOS output current (IDS) vs. input voltages (VGS and VDS) for various external and internal light spectrums will be presented. Techniques to optimize the internal quantum efficiency and photon absorption rate for various incident light frequencies will be illustrated. We will also present data for improved light sensitivity and changes in the output current due to the Photonic CMOS devices.

Finally, we will show the performance advantages of the Photonic CMOS for CMOS image sensor application - what are needed in the ROIC (Read Out Integrated Circuit) to process multiple multiplexed light signals.

Mid-infrared light sources are key components for future chemical sensing, on-chip optical interconnect and LIDAR technologies. GeSn alloys can be synthesized under silicon CMOS compatible conditions. Therefore, GeSn can open pathways for miniaturization of mid-infrared devices such as on-chip molecular sensors, optical interconnects and thermal cloaking devices. Most importantly, direct band gap GeSn is a promising material for on chip integration of light emitters. Previous research on GeSn films has shown mid-infrared lasing at increasing temperatures, initially (in 2015) at less than 90K and, in more recent reports, at or near room temperature. It has also been shown that GeSn nanowires with high-Sn content exhibit strong direct-gap photoluminescence at room temperature. Compared to GeSn thin films, GeSn nanowires are less constrained to lattice match with misfitting (e.g. Ge, Si) substrates and therefore may achieve higher Sn contents without strain and defects that promote non-radiative recombination. Additionally, unlike in a planar GeSn film where the emission angle is limited to < 13⁰ by its high refractive index (n>4), GeSn nanowires provide a platform to design the optical density of states for highly efficient light extraction directly from the light emitting material. GeSn nanowires may also constitute a superior medium for absorption and generation of photocarriers.
We demonstrate room temperature mid-infrared photodetection using resonantly absorbing GeSn/Ge core/shell nanowire photonic crystals. We have synthesized GeSn/Ge core/shell single crystal nanowires with 4% Sn that exhibit strong direct-bandgap photoluminescence at room temperature. The vertical nanowire photodetection device consists of arrays of nanowires in-filled with PMMA insulating layer, an ITO top contact and an aluminum bottom contact. Using full wave FDTD simulations, we optimize Mie resonances of individual nanowires (300 nm in diameter) at the desired bandgap wavelength. We further engineer the absorption with nanowires arranged in photonic crystal arrays. To synthesize the nanowire device, gold catalysts were patterned on Ge [111] substrate and then two step VLS growth was performed to synthesize Ge core and GeSn shell. We performed photocurrent characterization with an FTIR spectrometer at room temperature. The photocurrent spectrum of Ge nanowire photonic crystals indicates a four-fold enhancement due to resonant absorption, with the photocurrent spectrum tunable by varying the photonic crystal geometry.
For emission, we designed, synthesized and characterized mid-infrared emission of GeSn/Ge coreshell nanowires with 9-12% Sn. Full wave FDTD simulations reveal that the geometry of GeSn/Ge nanowires can be tuned to be resonantly scattering at the direct band gap energy. To further enhance light extraction, we use FDTD simulations to optimize GeSn/Ge nanowires in photonic crystal arrays to maximize the density of photonic states above the light line within the emission spectrum. The light emitting device consists of a similar device structure to that which is useful for photodetection. Room temperature photoluminescence characterization reveals both direct- and indirect-gap emissions in the mid-infrared, while electroluminescence characterization at 200 mA/cm reveals direct bandgap emission at both 2000 nm and 2300 nm. Optimization of nanophotonic light extraction efficiency will be discussed.

III-V compound semiconductor nanowires have been successfully used for a precise and simultaneous control of lattice parameters and bandgap structures bringing to existence a variety of functional nanoscale heterostructures and low-dimensional systems. Extending this paradigm to group IV semiconductors will be a true breakthrough that will pave the way to creating an entirely new class of silicon-compatible clean energy conversion, optoelectronic, and photonic devices. With this perspective, germanium-tin (GeSn) and silicon-germanium-tin (SiGeSn) alloys have recently been the subject of extensive investigations as new material systems to independently engineer lattice parameter and bandgap energy and directness.[1-5] The ability to incorporate Sn atoms into silicon and germanium at concentrations about one order of magnitude higher than the equilibrium solubility is at the core of these emerging potential technologies.
In this presentation, we will address the epitaxial growth and stability of these metastable nanowires with focus on Ge/GeSn core/shell nanowires with a sub-30 nm core. The reduced dimension of the Ge core was found to facilitate the growth of GeSn shell at a Sn content of ~10at.%, about 10-fold higher than the equilibrium solubility. We will discuss the optical and electronic properties and present strategies to integrate these nanowires in fabrication of short wavelength infrared (SWIR) and mid-infrared (MIR) optoelectronic devices. Finally, by using a microfabricated strain engineering platform, we will show that tensile strain allows tuning of the optical properties of Ge/GeSn core/shell nanowires in the SWIR and MIR range, thus laying the groundwork to implement innovative optoelectronic devices.

[1] A. Attiaoui, et al., J. Appl. Phys. 123, 223102 (2018)
[2] S. Assali, et al., Nano Lett 17, 1538−1544 (2017)
[3] M. Albani, S. Assali, et al., Nanoscale 10, 7250-7256 (2018)
[4] S. Assali, et al., Appl. Phys. Lett. 115, 113102 (2019)
[5] S. Assali, et al., arXiv:1906.11694

Sn-containing group IV semiconductors (Si)GeSn represent a versatile platform to implement a variety of Si-compatible photonic, optoelectronic, and photovoltaic devices. This class of semiconductors provides two degrees of freedom, strain and composition, to tailor the band structure and lattice parameter thus enabling a variety of heterostructures and low-dimensional systems on a Si substrate. In this presentation, we will discuss the recent progress in controlling and understanding the opto-electronic properties of metastable (Si)GeSn semiconductors. The relevance of these semiconductors for Si-compatible mid-infrared optoelectronics will be discussed form both materials and device perspectives. The growth of the (Si)GeSn multi-layer heterostructure with Sn contents up to ~20at.% is currently performed using a chemical vapor deposition (CVD) reactor on a Silicon wafer. By reducing the growth temperature, the Sn content in the alloy is increased, while preserving a high degree of crystal purity for the heterostructure in the topmost Sn-rich layer.[2,3] Atom probe tomography (APT) measurements will be discussed to address the abruptness of the interfaces in the GeSn multi-layer heterostructure and the uniform composition in the 18 at.% GeSn layer grown on top.[2] Positron annihilation lifetime spectroscopy (PAS) and depth-profiled Doppler broadening measurements will be discussed to enrich the understanding of point defects in these semiconductors. Based on these analyses, we found that divacancies are the predominant type of point defects in GeSn with composition in the 6.5-13 at.% range.[4]
Photoluminescence (PL) emission studies will be presented and discussed. In a GeSn layer with a Sn content of 18 at.% the room temperature PL emission was found to be centered at 0.36 eV (i.e. 3.5 μm wavelength).[2] However, the compressive in-plane strain (-1.3 %) in these GeSn layers reduces the directness of the alloy, leading to a higher energy gap value. By releasing the strain down to -0.2% in the 18 at.% Sn layer using a fully-underetched micro-disk geometry, a 50 meV red-shift of the PL emission energy down to 0.31 eV (i.e. 4.0 μm wavelength) is obtained. Moreover, the strained and relaxed PL emission and absorption measurements ranging from 300K down to 4K will be shown.[5] These observations will be discussed in the light of photocurrent measurements and photodetectors operating at MIR wavelengths.

[1] S. Wirths, et al., Prog. Cryst. Growth Charact. Mater. 62, 1 (2016)
[2] S. Assali, et al., Appl. Phys. Lett. 112, 251903 (2018).
[3] S. Assali, et al., J. Appl. Phys. 125, 025304 (2019).
[4] S. Assali, et al., Appl. Phys. Lett. 114, 251907 (2019)
[5] S. Assali, et al., submitted.

In a full-group IV integrated semiconductor platform for tensile-strained Ge the direct-band gap can be obtained when the Ge is grown on a lattice-mismatched Ge_0.87Sn_0.13 substrate. With this approach the main challenge is to increase the incorporation of Sn in Ge above the ~1 at.% equilibrium composition. Major developments were recently achieved in the epitaxial growth of random GeSn alloys with composition above 16 at.%. A biaxial tensile strain in a Ge layer up to ~1.5% was demonstrated when growing on a Ge_0.88Sn_0.12 substrate.[2-3] However, despite the large amount of tensile strain in Ge, no information is available on the abruptness of the Ge-GeSn interface and on the possibility of the subsequent GeSn growth on top.
Here, we discuss the epitaxial growth of a tensile-strained GeSn/Ge/GeSn (s-Ge) QW heterostructure with sharp interfaces and in-plane biaxial strain above 1.5 % grown on lattice-mismatched Ge_0.86Sn_0.14.[4] The sharpness of the Ge/GeSn heterostructure interfaces will be addressed using TEM-EELS and atom probe tomography (APT) measurements. A coherent s-Ge epitaxy with a tunable layer thickness in the sub-20 nm range is achieved. The pseudomorphic nature of the s-Ge layer will be discussed by combining cross-sectional STEM measurements and XRD-RSM measurements performed at a synchrotron radiation facility.
The s-Ge QW platform is a new versatile platform to investigate light-hole (LH) spin-based qubits and 2D high hole mobility electronics. Under tensile in-plane biaxial strain, LH occupies the top of the valence band and the LH-HH splitting can exceed 100s meV. Moreover, when the tensile in-plane biaxial strain approaches 2 % only LH are confined in the QW, thus minimizing LH-HH interactions
[1] R. A. Soref and L. Friedman, Superl. and Microstr. 14, 2/3 (1993).
[2] S. Wirths, et al., Appl. Phys. Lett. 102, 192103 (2013).
[3] V. A. Volodin, et al., JETP Letters 105, 5, 327–331 (2017).
[4] S. Assali, et al., submitted.

Integrated photonics have many compelling advantages for computing and communication applications, including high-speed and extremely wide bandwidth operations. When fused deposition printing (FDM) a PMMA plastic optical fiber on a surface the bottom of the fiber is flat due to non-uniform cooling and the force of gravity on the fiber during solidification. Here we show that by embedding the fiber in a urethane optical cladding material, the fiber cools homogeneously and retains its round ship which is optimum for light transmission. This homogeneous cooling allows the use of a flexible, low cost additive and subtractive assembly method called laser enhanced-direct print additive manufacturing (LE-DPAM) to produce plastic optical fiber interconnects for photonic systems we have enabled higher transmission rates, lower power requirements, improve signal integrity and timing, less heat generation, and improve security of communication signals.

Since the discovery of X-rays, the continuous expansion of X-ray technology has transformed our society. However, the capability gap for hard X-ray detectors in the 10-100 keV range is problematic. Even though scintillator-based methods are widely used, they are limited by their decay time response, light yield, and spatial resolution, which may prevent them from being effective in the next generation of light source facilities, the generic platform of imaging sensor technology, and quantum image computing.

One class of X-ray detection primarily focuses structured photocathode designs that rely on external photoemission and collecting electrons converted from photons using an external electric field. This research only demonstrated up to 5% photon-to-electron conversion efficiency (electron generation) at 7.5keV incident photon energy, which is still commendable as this value has been roaming around the 1-1.5% area for external photoemission for a few decades [1, 2]. However, we believe that there is a potential to significantly increase this 5% electron generation even beyond 7.5keV incident photon energy via X-ray photon emission from high-Z materials.

Here, we present the Monte Carlo simulation results and analyses via Monte Carlo N-Particle Software (MCNP) that demonstrate a tremendous potential of revolutionizing the advancement of X-ray imaging with a two-layer design that consists of a high-Z semiconductor material on top of a Si layer. The high-Z layer is called photon attenuation layer (PAL), and the Si layer the electron detection layer (EDL). PAL-EDL latticed multi-pixel X-ray detector can also lead to low inter-pixel scattering with significant electron and photoelectron generation. The underlying principle of this design is photon energy down conversion, where the top high-Z PAL leads to an effective X-ray photon energy attenuation down to a few keV and allows Si to absorb them with a much higher absorption coefficient. This approach is distinct from scintillators in that the attenuated photons remain in X-ray spectral regime instead of visible regime. For instance, with PbTe, CdTe or CZT PAL and the Si EDL, we have observed >22% electron generation at 10keV X-ray photon energy and >2.8% electron generation at 40 keV incident photon energy with minimum inter-pixel scattering effects. The average electron energy lies between a few hundred eV and a couple keV.

At 30-40keV, we observed that adding a high-Z PAL on top of Si EDL can lead up to 13x higher electron generation than the typical Si direct-detection method seen in commercial products such as photomultipliers. Furthermore, because MCNP does not include the “avalanche gain” process in Si, the observed efficiencies may be underestimated. The MCNP results will be coupled with models that take the avalanche gain process into consideration (e.g. TCAD). In addition, this high-Z PAL-Si EDL configuration can be integrated to regular backside illumination (BSI) CMOS image sensor or quanta image sensor-based devices, paving the way towards a billion-pixel X-ray camera design for wide field-of-view applications in light sources [3].

[1] Y.P. Opachich et al., "Structured photocathodes for improved high-energy x-ray efficiency in streak cameras, Review of Scientific Instruments", vol. 87, no. 11, pp. 11E331, 2016.
[2] B. Henke, J. Knauer and K. Premaratne, "The characterization of x-ray photocathodes in the 0.1-10keV photon energy region", Journal of Applied Physics, vol. 52, no. 3, pp. 1509-1520, 1981.
[3] Z. Wang, C. W. Barnes, D. M. Dattelbaum, E. R. Fossum, E. Lee, J. Liu, J. J. Ma, R. Pokharel, Y. H. Sechrest, C. M. Sweeney, and A. C. Therrien, "A billion-pixel camera (BiPC) for wide field-of-view measurements in synchrotrons and X-ray Free Electron Lasers", Los Alamos Report, LA-UR-19-27804, 2019.

This work presents the first observation of space charge limited conduction (SCLC) mechanism in intrinsic GaAsSb nanowires (NWs) grown by Ga-assisted molecular beam epitaxy and the effect of post-growth in-situ annealing in an ultra-high vacuum on trap concentration and its energy distribution in the NWs. Current-voltage (I-V) measurements on single NW (using conductive atomic force microscopy) and ensemble NWs (using two probe method) exhibited linear behavior at lower bias transitioning to a power-law behavior at higher bias, where the dominance of injected carriers over thermally generated charge carriers was observed. Temperature-dependent I-V analysis (I-V-T) on as-grown ensemble NW device in SCLC region yielded a trap concentration of 10¹⁶ cm^-3 distributed over a wide energy distribution in the bandgap as opposed to the reduced trap concentration of 7 * 10¹⁴ cm^-3 in in-situ annealed NW ensemble confined to a narrower energy distribution of 0.12 eV located below the band edge, suggesting that annealing in ultra-high vacuum is an effective approach for the annihilation of the traps. The trap energy level in in-situ annealed NWs is speculated to be originating from Ga vacancy, and Ga_Sb anti-site defect level during growth. Observations of increased PL intensity with reduced full-width half maxima at 4K and lower LO mode in corresponding Raman spectra for in-situ annealed NWs compared to as-grown NWs further attest to the annihilation of traps on in-situ annealing. Further, AFM current mapping also exhibited a considerable increase in several conducting NWs in in-situ compared to as-grown NWs. Hence, I-V-T analysis of the SCLC mechanism has been demonstrated to be a simple approach to obtain information on growth induced traps in the NWs.

Photon-generated carriers may substantially improve the output current, and reduce series resistance of CMOSFETs. Photonic devices, such as Multiple Quantum Well Lasers, Quantum Dots Lasers, Vertical Cavity Surface-Emitting Lasers (VCSEL), Light Emitting Diodes (LEDs) have been reported for various applications. With modulation doped laser junctions, and bandgap engineered tunneling mechanisms, ultra-low resistivity and laser diode forward voltage can be achieved. These laser diodes can be fabricated in the drain region of a MOSFET. A photon sensor, such as avalanche photo diode (APD), can be included in the channel / well regions. The laser, APD, and MOSFET are fabricated as one integral transistor. When the MOSFET is turned off, the laser is also turned off. The laser is switched on only when a gate voltage and a drain voltage are applied to the MOSFET. Light emitted from the laser is absorbed by the APD. Light current produced by the built-in photon sensors or APD reduces series resistance and may substantially improve the output current and switching speed.

Indium Tin Oxide (ITO) is commonly used in Charge Coupled Devices (CCD) receiving front illumination. The gate electrode material can be replaced with ITO, which is transparent for specific light spectrums. The new Photonic CMOS transistor, including nonlinear optical films and multiple photon sensors (for various light spectrums) can be used for CCD (with ITO) or CMOS image products (without ITO) for front or back illumination.

When external light is absorbed in the depleted channel and well regions (where the multiple photon sensors or APDs are located) of a Photonic MOSFET, light electrons are generated. These light electrons produce a drain current, which turns on the built-in laser in the drain region. The result is not only a light current, but also a new beam of a laser light are generated. Much higher resolution may be achieved with multiplexing of the light signals and specially designed ROIC (Read Out Integrated Circuit), which may also include Photonic CMOSFETs.

Nonlinear Optics have been reported for various products and applications. The nonlinear crystal films change the input light signal and produces an output light signal of a different frequency or color. In the new Photonic CMOS transistor, in addition to the built-in laser in the drain, there are multiple avalanche photo diodes in the channel / well regions, nonlinear crystal films below the APDs, and a reflector (metal or multiple indices total reflection film stack) in the bottom of the device. The APDs are designed to absorb lights from various bands of spectrums. When light passes through the Photonic CMOS Transistor, some APD films absorb lights of specific frequencies, and other APD films absorb the reflected light of different frequencies through the nonlinear crystal layers. This feature enhances the absorption rate and external quantum efficiency of the Photonic MOSFET, and may significantly improves the quality and resolution of the images from the CCD or CMOS imaging devices.

In this paper we will discuss Photonic MOSFET drain current (IDS) vs. spectral characteristics of nonlinear optical films. With the nonlinear optical films, it is possible to substantially improve the CMOS performance according to these data. Dark current density and resolution using a Photonic CMOS image sensor vs. traditional CMOS image sensor are analyzed. The advantages of a Photonic CMOS image sensor are lower dark current, and higher light sensitivity, due to the built-in lasing device and multiple photon sensing films, and nonlinear optical features.

Information regarding incident light spectrums vs. Ion / Ioff ratio for a Photonic CMOS image sensor will presented. Comparison of the speeds of CMOS image sensors with and without the built-in nonlinear optics films will be discussed.

We will also introduce various types of crystals of nonlinear optics suitable for the Photonic CMOS image sensing applications.

Owing to their large oscillator strength and strong exciton binding energy, two-dimensional (2D) transition metal dichalcogenides (TMDs) have emerged as an attractive material platform for studying strong light-matter coupling and associated phenomena. Following up on our previous work on strong light-matter coupling in 2D TMDs [1, 2], here we will present results on enhancing the nonlinear interaction between the quasiparticles (exciton-polaritons). This is achieved using excited states of excitons (Rydberg states) which have larger Bohr radii. We will also present our recent results on controlling the valley pseudospin via the pseudomagnetic fields in optical cavities and our work on realizing an electrically pumped polariton LED [3]. Finally, we will first present our work on realizing single photon emitters (SPEs) in hexagonal boron nitride (hBN), a van der Waals material, via strain engineering [4] and coupling of these SPEs to high Q silicon nitride microresonators [5].
[1] X. Liu, et al., Nature Photonics 9, 30 (2015)
[2] Z. Sun et al., Nature Photonics 11, 491 (2017)
[3] J. Gu et al. Nature Nanotech. (2019) DOI: 10.1038/s41565-019-0543-6
[4] N. Proscia et al. Optica 5, 1128 (2018)
[5] N. Proscia et al. ArXiv 1906.06546

One of very important requirements for nanolasers intended for future photonic chips is the energy consumption. The amount of total energy input to a laser required to a large extend is determined by the critical value of carrier density, at which optical gain occurs. In conventional semiconductors, optical gain is based on electron-hole plasma and it occurs at the so-called transparency density, which is near and above the Mott density, typically on the order of 10¹² per square centimeter. In 2D materials, the orders of magnitude larger exciton binding energy make it possible to have optical gain based on various excitonic species such as excitons, trions, bi-excitons etc. Such optical gain can occur at density levels several orders of magnitude smaller, well below the Mott density. In this talk, we will discuss possibilities of such optical gain in 2D materials that could occur at extremely low injection levels. As an example, we will discuss our recent experimental results [1] on trion gain in a gated 2D MoTe2 structure that occurs at the density level ~5 orders of magnitude smaller than the Mott density.

References:
[1] Z. Wang, H. Sun, Q. Zhang, J. Feng, J. Zhang, Y. Li, C.Z. Ning, Excitonic Complexes and Optical Gain in Two-Dimensional Molybdenum Ditelluride Well below Mott Transition, arXiv:1812.04296 [cond-mat.mes-hall]

With ever-increasing demand for faster speeds, copper interconnects that ferry vast amounts of data in computing devices are expected to significantly lag behind the requirements, thereby creating a bottleneck for moving data rapidly. Optical interconnects provide a promising alternative, allowing much faster speeds and larger bandwidths between different components, on or off chip. Three most critical interconnect components are light sources, waveguides and detectors. Currently, optical communication is performed by a combination of these devices where the information is encoded in photon number and frequency (intensity and bandwidth). However, no information is encoded in intrinsic properties of photons, i.e., photon spin and also possibly the orbital angular momentum modes of the light beam. Analogous to the idea of spintronics, where the spin degree of freedom of electrons are utilized to encode and transmit information, similar enhancements in the data processing capabilities of the optical interconnect systems can be obtained if information can be encoded in photon spin. Photons, have two fundamental spin polarizations reflected in their polarization state, i.e., right and left circular polarization. To fully utilize more degrees of freedom of light to enhance the data processing capabilities of network systems, it is required that new light sources, detectors and waveguides be designed and fabricated that can produce, transmit and detect light with complex polarizations. We will discuss some recent developments in our laboratory towards the development of topological waveguides that can route photonic signals based on photon spin and also photodetectors that are sensitive to circular polarization (photon spin). Either by protecting or breaking certain symmetries in (meta)materials, new photonic materials and devices will be discussed that can enable a new generation of photonic devices that can encode, transmit and sense information encoded in spin and angular momentum modes of light that are promising for the development of integrated chiral photonic systems with extremely large data processing capabilities.

We introduce a physical mechanism to perform machine learning by demonstrating a Diffractive Deep Neural Network (D²NN) architecture that can all-optically implement various functions following the deep learning-based design of passive layers that work collectively. We created 3D-printed diffractive networks that implement classification of images of handwritten digits and fashion products as well as the function of an imaging lens, spectral filters and wavelength demultiplexers at terahertz spectrum. This passive diffractive network can perform, at the speed of light, various complex functions that computer-based neural networks can implement, and will find applications in all-optical image analysis, feature detection and object classification, also enabling new camera designs and optical components that perform unique tasks using diffractive neural networks designed by deep learning.

Insects are capable of amazing tasks, such as traveling across hundreds of kilometers of unfamiliar terrain to pinpoint a specific breeding ground, with a brain that contains 100 000x fewer neurons than our own brains, using merely a few drops of nectar as energy supply. At each moment during navigation, the insect has to decide whether to turn right, left or go straight. Integrating sensory inputs, internal states, and previous experience, it computes its current and intended headings, compares the two, and initiates steering. A brain region called the central complex (CX) is essential for these tasks. Recently a biologically constrained computational model of the CX has been generated that can carry out these computations to generate homing behavior [1]. This computational model has been shown to carry out homing tasks using speed and heading information of limited precision, with considerable noise in the circuit. Containing less than 100 neurons, it can serve as a model system to understand the functionality and prove the efficiency of novel hardware solutions.

To realize this CX model we propose and theoretically model a network circuit based on existing III-V nanowire (NW) opto-electronic components that are optically connected through a shared waveguide structure. The neural weights of the mutual interconnections are achieved by controlling the far-field light emission patterns and by strategic placement of the nano-components. As a result, there is no need for physical connecting wires between the network nodes. Each node in this network is capable of both receiving, evaluating and sending optical signals, while power consumption is minimized by using high efficiency concepts from NW based photovoltaics and quantum dot emitters [2-6]. We show that the modeled coupling efficiencies and activation functions well suffice to facilitate excellent navigation in the insect neuron model [1]. As the efficiencies and electrical performance of the NW devices are known, we can estimate the energy consumption per operation and find that this should be at least as low as in biological processes [7] and beyond standard hardware implementations.

Our detailed simulations of the fundamental device component, which will act as a basic sigmoid neural node throughout the network, show how it performs the tasks of receiving light inputs, evaluating them and emitting a light output. We model it as an electrical circuit which is parametrized and solved by a standard SPICE solver (Ngspice). We describe the subcomponents using an Ebers-Moll model and the Shockley diode equation, respectively, where the relevant properties were calculated using a drift-diffusion model with thermionic boundary conditions in COMSOL. A NW phototransistor was optimized to give enough gain to broadcast the signal to the next layer using a realistic non-radiative Shockley-Read-Hall recombination lifetime for NWs of 1.34 ns [8] as well as Lorenz-Mie scattering theory. Extensive finite difference time domain (FDTD) simulations (Lumerical), were made for the central part of the CX model which consists of eight fully interconnected components. They were placed in a shared asymmetric waveguide which was optimized for the target wavelength. As the neurons of this layer are interconnected, input-output isolation is non-trivial, and the FDTD model was used to optimize the network in this respect. The gain of the fundamental device ensures both cascadeability and proper fan-out without the need of critical biasing [9], and the analog architecture carries a built-in robustness to noise [1].
[1] T. Stone, et al, Curr. Biol. 27, 3069 (2017)
[2] J. Claudon, et al, Nat. Photonics 4, 174 (2010)
[3] M. E. Reimer, et al, Nat. Commun. 3, 737 (2012)
[4] J. Wallentin, et al, Science 339, 1057 (2013)
[6] G. Otnes and M. T. Borgström, Nano Today 12, 31 (2017).
[7] C. Mead, Proc. IEEE 78, 1629 (1990)
[8] H. J. Joyce, et al, Nano Lett. 12, 5325 (2012)
[9] D. A. B. Miller, Nat. Photonics 4, 3 (2010)

Optical analog computing provide solutions with fast and power-efficient performance for big-data applications. Recently investigated metamaterials and metasurfaces demonstrate potentials in mathematical operations over wavelength-scale footprint. We extend the functionality of using nonlocal dielectric metasurface to perform spatial differentiation equation for imaging processing. The study pave the way towards ultrathin devices for optical signal processing.

Superconducting qubit quantum computers offer new capabilities for quantum information processing, but require < 50 mK environments and near complete isolation from outside noise sources. To enable quantum information to be exchanged with these systems outside of cryogenic conditions by means of higher energy photons, many approaches towards microwave-optical transduction are being explored by the research community. Three-wave mixing in linear electro-optic materials is a particularly appealing approach due to the inherent simplicity and elimination of excess noise sources that are introduced by intermediate states present in many alternative technologies. However, material microwave and optical loss rates have limited conversion efficiency below 50% in demonstrations to date. Here, I present IBM’s work towards integrating low loss SiGe/Si optical resonators with the low loss Nb-on-Si microwave resonators that have been developed for superconducting qubits. An effective linear electro-optic coefficient for coupling the microwave and optical fields by three wave mixing is induced by strongly biasing the microwave resonators to leverage the DC Kerr effect in Si and SiGe. For 1 MHz-rate transducers, the resulting microwave-optical coupling strength is calculated to be sufficient to reach the peak conversion efficiency condition. Demonstrating microwave-optical transduction in this platform is then dependent on achieving low defectivity fabrication of the proposed SiGe/Si optical waveguides. I will present a study of this materials problem and an analysis of the overall optimal transducer design.

This work was funded by LPS/ARO under CQTS program, contract number W911NF-18-1-0022.

Manipulation of single photons generated deterministically from pre-specified sources on-chip to create multiphoton entangled states remains a major goal to be reached in photonic quantum information
processing platforms. Notable advances have been made in single photon source performance and
demonstrations of photon interference and entanglement based largely upon finding suitable single photon
source from an otherwise random ensemble. The lack of scalable on-chip architectures stands as a major
obstacle to realizing compact quantum information technologies.
In this talk I will present our proposition [1,2] and continuing efforts [3,4] to realize such a
photonic chip built upon a unique class of spatially-ordered arrays of surface-curvature driven mesa-top
single quantum dots (MTSQDs) [1,3]. These AlGaAs/InGaAs MTSQDs exhibit spectral uniformity as
low as 1.8nm across 5×8 arrays, exhibit pairs of as-grown QDs having emission energy within 200μeV,
and single photon purity > 99%. Following a planarization overlayer growth, these buried single photon
sources (SPS) provide the essential platform for subsequent lithographic fabrication of emitted photon
light manipulating units (LMUs). The LMUs exploit either the well-developed photonic 2D crystal
approach or a new approach we introduced [1,2] based upon Mie-like resonances in interacting
subwavelength-sized dielectric building block (DBB) based metastructures. The in-plane single Mie-like
resonance of the metastructure (the LMU) provides all the needed basic photon manipulation functions to
create controlled on-chip interference and entanglement [4]: enhancement of the SPS emission rate
(Purcell enhancement), directional emission (local antenna), state-preserving propagation, path
bifurcation (beam-splitting), and beam combining. Findings of our combined theoretical analysis,
numerical simulations, and experimental studies of such SPS-LMU primitives — the essential building
unit for quantum optical circuits — will be presented. The emphasis is on the holistic approach to proofof-
principle demonstration of the needed on-chip multifunctional system with its built-in trade-offs rather
than the best of any individual light manipulation function.
The work is supported by the US Army Research Office (W911NF1910025) and the Air Force Office of
Scientific Research (FA9550-17-1-0353).
1. Jiefei Zhang et al., Jour. App. Phys. 120, 243103 (2016)
2. Swarnabha Chattaraj et al., Jour. Opt. Soc. America B. 33, 2414(2016)
3. Jiefei Zhang, et al., App. Phys. Lett 114, 071102 (2019)
4. Swarnabha Chattaraj et al., IEEE Jour. Quant. Elec. 56, 1 (2019)

In recent years, InAs quantum dot lasers have proven themselves resilient against crystalline defects that form during heteroepitaxial growth of III-Vs on silicon. The three-dimensional confinement and reduced in-plane diffusion lengths of carriers in quantum dots relative to quantum wells are the source of this resiliency, but alone these characteristics are not sufficient for demonstrating commercially reliable performance for lasers on silicon. Quantum dot active regions must be utilized together with highly optimized III-V/Si buffers to reduce the dislocation density to achieve acceptable device reliability.
In this talk, recent advances in GaAs/Si templates will be highlighted along with the resulting enhancements in laser reliability. Over the last three years, we have produced InAs quantum dot lasers on GaAs/Si templates with dislocation densities ranging from 3×10⁸ cm^-2 down to < 7×10⁶ cm^-2 by optimizing thermal cycling procedures and strained filter layers. As the material has improved, the extrapolated mean-time-to-failure has increased several orders of magnitude to >10,000,000 h at 35°C and >50,000 h at 60°C.
In addition to improving laser reliability, the improvements in material quality have virtually closed the performance gap relative to native substrate lasers before aging. At dislocation densities of 7×10⁶ cm^-2, low threshold, high power, and high efficiency lasers are readily achievable with continuous wave operation above 100°C. These improvements justified more advanced laser implementations and investigations into the dynamic performance of quantum dot lasers on silicon. The unique atom-like density of states of quantum dots has historically enabled unique properties such as an insensitivity to optical feedback, highly engineerable gain bandwidth, and ultrafast gain recovery; these properties unique to quantum dots have enabled Fabry-Perot lasers with complete insensitivity to optical feedback, high performance mode-locked lasers operating at 20 GHz with 64 channels capable of 4.1 Tbps transmission below the forward error correction (FEC) limit, and 100 GHz colliding pulse mode-locked lasers capable of 0.9 Tbps transmission below the FEC limit. Achieving each of these results required a high degree of optimization in the epi design and materials growth to obtain uniform distributions of high-quality quantum dots, which will be emphasized in the talk.
Further improving performance and, in particular, demonstrating commercially viable reliability at elevated temperatures >60°C will require additional efforts to reduce the dislocation density in the laser active region. Our efforts to push the threading dislocation density < 2×10⁶ cm^-2 while also reducing the buffer thickness will be discussed in addition to efforts to address misfit dislocations that have been observed in the laser active region. Given the much larger cross-section of material that interacts with a misfit dislocation relative to a thread, it is believed that addressing these defects is critical to reaching the goal of native substrate level device reliability on a silicon substrate.

Mid-infrared light sources are key components for future chemical sensing, on-chip optical interconnect and LIDAR technologies. GeSn alloys can be synthesized under silicon CMOS compatible conditions. Therefore, GeSn can open pathways for miniaturization of mid-infrared devices such as on-chip molecular sensors, optical interconnects and thermal cloaking devices. Most importantly, direct band gap GeSn is a promising material for on chip integration of light emitters. Previous research on GeSn films has shown mid-infrared lasing at increasing temperatures, initially (in 2015) at less than 90K and, in more recent reports, at or near room temperature. It has also been shown that GeSn nanowires with high-Sn content exhibit strong direct-gap photoluminescence at room temperature. Compared to GeSn thin films, GeSn nanowires are less constrained to lattice match with misfitting (e.g. Ge, Si) substrates and therefore may achieve higher Sn contents without strain and defects that promote non-radiative recombination. Additionally, unlike in a planar GeSn film where the emission angle is limited to < 13⁰ by its high refractive index (n>4), GeSn nanowires provide a platform to design the optical density of states for highly efficient light extraction directly from the light emitting material. GeSn nanowires may also constitute a superior medium for absorption and generation of photocarriers.
We demonstrate room temperature mid-infrared photodetection using resonantly absorbing GeSn/Ge core/shell nanowire photonic crystals. We have synthesized GeSn/Ge core/shell single crystal nanowires with 4% Sn that exhibit strong direct-bandgap photoluminescence at room temperature. The vertical nanowire photodetection device consists of arrays of nanowires in-filled with PMMA insulating layer, an ITO top contact and an aluminum bottom contact. Using full wave FDTD simulations, we optimize Mie resonances of individual nanowires (300 nm in diameter) at the desired bandgap wavelength. We further engineer the absorption with nanowires arranged in photonic crystal arrays. To synthesize the nanowire device, gold catalysts were patterned on Ge [111] substrate and then two step VLS growth was performed to synthesize Ge core and GeSn shell. We performed photocurrent characterization with an FTIR spectrometer at room temperature. The photocurrent spectrum of Ge nanowire photonic crystals indicates a four-fold enhancement due to resonant absorption, with the photocurrent spectrum tunable by varying the photonic crystal geometry.
For emission, we designed, synthesized and characterized mid-infrared emission of GeSn/Ge coreshell nanowires with 9-12% Sn. Full wave FDTD simulations reveal that the geometry of GeSn/Ge nanowires can be tuned to be resonantly scattering at the direct band gap energy. To further enhance light extraction, we use FDTD simulations to optimize GeSn/Ge nanowires in photonic crystal arrays to maximize the density of photonic states above the light line within the emission spectrum. The light emitting device consists of a similar device structure to that which is useful for photodetection. Room temperature photoluminescence characterization reveals both direct- and indirect-gap emissions in the mid-infrared, while electroluminescence characterization at 200 mA/cm reveals direct bandgap emission at both 2000 nm and 2300 nm. Optimization of nanophotonic light extraction efficiency will be discussed.

Artificial Intelligence (AI) is transforming our lives in the same way as the advent of the Internet and cellular phones has done. AI is revolutionizing the healthcare industry with complex medical data analysis, actualizing self-driving cars, and beating humans at strategy games such as Go. However, it takes thousands of CPUs and GPUs, and many weeks to train the neural networks in AI hardware. Traditional CPUs, GPUs, and neuromorphic (i.e., brain-inspired) electronics such as the IBM TrueNorth and Google TPU, may find it a challenge to train the neural networks of the future.

Neuromorphic photonics has experienced a recent surge of interest over the last few years, promising orders of magnitude improvements in both speed and energy efficiency over digital electronics using: artificial neural networks, spiking neural networks, and reservoir computing. By combining the high bandwidth and efficiency of photonic devices with the adaptive, parallelism and complexity attained by methods similar to those seen in the brain, photonic processors have the potential to be at least ten thousand times faster than state-of-the-art electronic processors while consuming less energy/computation.

We will provide an overview of neuromorphic photonic systems and their application to optimization and machine learning problems. We will discuss the physical advantages of photonic processing systems and describe underlying device models that allow practical systems to be constructed. We also describe several real-world applications for control and deep learning inference. Lastly, we will discuss scalability in the context of designing a full-scale neuromorphic photonic processing system, considering aspects such as signal integrity, noise, and hardware fabrication platforms. This talk is intended for a wide audience and hopes to teach how theory, research, and device concepts from neuromorphic photonics could be applied in practical machine learning systems.

Integrated photonic circuits efficiently combine multiple optical functions on a single physical platform, such as guiding, modulating, splitting, or detection of light. The integration of these functions removes barriers for designing and realizing large optical circuits, as present when using discrete components like bulky lenses or mirrors in tabletop optical networks. With the rise of silicon photonics, integrated photonic circuits (PICs) are becoming increasingly large and highly functional, eventually allowing unprecedented concepts of photonic computing. Examples are integrated photonic quantum processors, microwave photonic filters and processors, and optical accelerators to train and execute neural networks.

An important building block in integrated optical circuits is an efficient link between the optical and electrical domain. Well-known examples of such links are integrated high-speed modulators to convert electrical signals into optical signals at very high-speed, and low-power tuning elements to compensate for variations in the device operation temperature and for device-to-device variations during fabrication. To enable such electro-optic links, the two most widely used physical effects are the plasma-dispersion effect and Joule heating. Although these effects are attractive to use due to their compatibility with standard photonic fabrication processes, their performance in integrated devices is intrinsically limited by high insertion losses and high-power dissipation.

Over the past decade, we established an alternative electro-optic switching technology by embedding a Pockels material into silicon-based photonic devices [1]. We reached this goal by developing a process to fabricate ferroelectric barium-titanate (BTO) thin films on silicon substrates using advanced epitaxial deposition techniques and by developing a BTO process technology [2]. We correlated the electro-optical properties of the thin films with their structural properties such as porosity and crystalline symmetry to show guidelines for improving the functional properties. By realizing integrated hybrid BTO/silicon devices, we demonstrated record-high, in-device Pockels coefficients of >900 pm/V. The Pockels effect in BTO-based photonic devices indeed enables extremely fast data modulation at rates beyond >40 Gbps and ultra-low-power electro-optic tuning of silicon and silicon-nitride waveguides. We also show ways of how to integrate and use BTO in plasmonic slot waveguide structures for very compact optical devices. With the development of a wafer-level integration scheme of single-crystalline BTO layers to a 200 mm process [3], we could demonstrate a viable path to combine the BTO-technology with existing fabrication routes.

With major breakthroughs in the past years, BTO has emerged as a strong candidate for a novel generation of electro-optic devices. Major achievements of the BTO technology will be covered in the presentation, ranging from important materials aspects, device development, integration concepts, and novel applications in the area of quantum computing, high-speed communication, and neuromorphic optical computing.

[1] S. Abel et al., “Large Pockels effect in micro- and nano- structured barium titanate integrated on silicon,” Nat. Mater., vol. 18, no. 1, pp. 42–47, 2019.
[2] S. Abel et al., “A strong electro-optically active lead-free ferroelectric integrated on silicon,” Nat. Commun., vol. 4, p. 1671, 2013.
[3] F. Eltes et al., “A BaTiO3-Based Electro-Optic Pockels Modulator Monolithically Integrated on an Advanced Silicon Photonics Platform,” J. Light. Technol., vol. 37, no. 5, pp. 1456–1462, 2019.

The neutral divacancy (VV0) in silicon carbide (SiC) exhibits robust spin coherence and a high-quality near-infrared spin-photon interface in a material compatible with mature fabrication techniques. Here, we make use of this scalable semiconductor host and design electronic devices to manipulate embedded isolated quantum systems. Specifically, we create and isolate single spin defects in a commercial p-i-n diode [1]. This simple integration enables engineering of the defect’s charge environment and drastically reduces electric field noise. Surprisingly, the use of electrical gating mitigates spectral diffusion and achieves near-lifetime limited optical linewidths. Furthermore, by exploiting field confinement of the junction, we show that the optical transitions can be gate-tuned by nearly a terahertz, revealing a spectral tuning range extending over 40,000 optical linewidths. This geometry may allow for spectral multiplexing of many quantum channels, and also provides a method for using electric fields combined with optical excitation for deterministic charge state control.

Applying gigahertz ac electric fields to SiC devices produces coherent interference in the form of Landau-Zener-Stückelberg fringes, arising from interactions between microwave and optical photons [2], even in the absence of a microwave resonator. We demonstrate lifetime-limited optical coherence and clock-like spin transitions with increased robustness against magnetic noise. Electrical driving of excited-state electron orbitals offers advantages over spin-based coupling and points towards new types of hybrid quantum systems. These results reveal new opportunities for electrical manipulation of spin-based quantum systems in scalable SiC electronic devices.

This work was done in collaboration with S. L. Bayliss, A. L. Crook, P. J. Mintun, S. J. Whiteley, G. Wolfowicz, H. Abe, A. Gali, V. Ivady, T. Ohshima, G. Thiering, P. Udvarhelyi.

[1] C. P. Anderson*, A. Bourassa*, K. C. Miao, G. Wolfowicz, P. J. Mintun, A. L. Crook, H. Abe, J. U. Hassan, N. T. Son, T. Oshima, D. D. Awschalom, “Electrical and optical control of single spins integrated in scalable semiconductor devices,” Science, accepted for publication (2019); arXiv: 1906.08328

[2] K. C. Miao, A. Bourassa, C. P. Anderson, S. J. Whiteley, A. L. Crook, S. L. Bayliss, G. Wolfowicz, G. Thiering, P. Udvarhelyi, V. Ivady, H. Abe, T. Ohshima, A. Gali, D. D. Awschalom, “Electrically driven optical interferometry with spins in silicon carbide,” Science Advances, in press (2019); arXiv: 1905.12780

Semiconductor materials play a pivotal role in photonic technology. They can be used as tailored photon emitters and detectors. Combined with plasmonic structures and photonic design, their efficiency and functionality is further extended.
In this talk we show how compound semiconductor nanostructures such as nanowires and nananoscale membranes (nanofins) can be used for both single photon emission and detection [[1],[2]]. We will explore what kind of heterostructures are optimal for these applications. We will also elucidate how photon detection can be optimized by the nanowire geometry and the combination with plasmonic structures [2,[3]]. Finally, we will also explore materials with a lower production and environmental cost and how to fabricate them at a similar functionality than compound semiconductors. We will show how GeSn thin films and nanostructures have the potential to become the leading material in mid-infrared photodetection.
[1] M. Heiss et al Nature Mater. 12, 439 (2013)
[2] A. Dorodnyy et al IEEE J. Sel. Top. Quant. Electr. 24, 1-13 (2018)
[3] A. Casadei et al, Sci. Rep. 5, 7651 (2015)

Combinatorial optimization problems are ubiquitous in our modern life. Classical examples include lead optimization in drug and biocatalyst discovery, resource optimization in wireless communications, logistics and scheduling, sparse coding for compressed sensing, deep machine learning in artificial intelligence and fintech. These optimization problems can be mapped to either Ising model, XY model or k-SAT problem, which is a main reason why various Ising, XY and SAT solvers have been proposed and implemented in the recent years.

We have focused on the network of optical parametric oscillators to construct coherent Ising machines, XY machines and SAT solvers. An optical parametric oscillator (OPO) operates as a quantum analog device at below oscillation threshold and also functions as a classical digital device at above oscillation threshold. We need not only quantum computational resources, such as quantum correlation and quantum suppression of chaos, but also classical computational resources, such as spontaneous symmetry breaking and exponential amplitude amplification, to build an efficient optimizer. An OPO is almost a unique choice of device to realize such quantum and classical computational resources simultaneously at room temperatures.

A study on quantum state engineering in an OPO with measurement-feedback control was dated back to 1980s [1,2]. A PPLN waveguide OPO device was identified as a stable and efficient generator for squeezed vacuum state pulses at communication wavelengths in 1990s [3]. An idea of using the pitchfork bifurcation and spontaneous symmetry breaking at OPO threshold as an irreversible decision making process was proposed in 2013 [4] and experimentally implemented [5,6]. A scalable architecture based on measurement-feedback coupling scheme was subsequently demonstrated [7,8]. Those devices already demonstrated competitive performance against modern heuristics [9] and quantum annealers [10]. In this talk, we will discuss the principles of quantum-classical crossover and future prospects of quantum neural network.

[1] G. Bjork and Y. Yamamoto, Phys. Rev. A 37, 4229-4239 (June 1988).
[2] K. Watanabe and Y. Yamamoto, Phys. Rev. A 38, 3556-3565 (October 1988).
[3] D. K. Serkland et al., Opt. Lett. 20, 1649-1651 (August 1995).
[4] Z. Wang et al., Phys. Rev. A 88, 063853 (December 2013).
[5] A. Marandi et al., Nature Photonics 8, 937-942 (October 2014).
[6] T. Inagaki et al., Nature Photonics 10, 415-419 (June 2016).
[7] P. L. McMahon et al., Science 354, 614-617 (October 2016).
[8] T. Inagaki et al., Science 354, 603-606 (October 2016).
[9] T. Leleu et al., Phys. Rev. Lett. 122, 040607 (February 2019).
[10] R. Hamerly, Science Advances 5, eaau0823 (May 2019).