Symposium EP08—Phase-Change Materials for Memories, Photonics, Neuromorphic and Emerging Application

On-demand phase transformation of silicon is of great technical importance for electronics, MEMES and photonics. While numerous efforts have been made towards understanding the phase change of thin films, scientific knowledge is limited in nanoscale phase transformation.

Here we probed the nanosecond resolved crystallization and amorphization dynamics of a single silicon nanostructure with 200nm diameter. Based on the Mie resonance and thermo-optic effect, we successfully identified the heating, melting, deformation, solidification and cooling states during the phase transformations. The results revealed the super-lateral crystallization of 2ns time scale and amorphization of 10ns. Combined with comprehensive heat-transfer and Monte Carlo based nucleation and phase change simulation, the study offered significant insights into the different super-coolings, cooling rates and corresponded mechanisms of crystallization and amorphization.<embed height="0" id="xunlei_com_thunder_helper_plugin_d462f475-c18e-46be-bd10-327458d045bd" type="application/thunder_download_plugin" width="0" />

Materials with rationally controlled properties play important parts in the development of new and advanced technologies. For instance, the properties of thermoelectric, phase-change, or topologically insulating materials can be traced back, to a significant extent, to the nature of bonding in materials. Here, we develop a two-dimensional map based on a quantum-topological description of electron sharing and electron transfer. This map intuitively identifies the fundamental nature of ionic, metallic, and covalent bonding in a range of elements and binary materials. Furthermore, it highlights a distinct region for a mechanism recently termed “metavalent” bonding. Extending this map into the third dimension by including physical properties of application interest, we show that bonding in metavalent compounds differs from the classical textbooks views of bonding. This map could be used to help designing new materials: by searching for desired properties in a 3D space and then mapping this back onto the 2D plane of bonding.

It is almost ten years since interfacial phase-change memory (iPCM) was developed. Although iPCM was originally designed to save switching energy of phase-change memory (PCM) by reducing entropy energy loss at the phase transition as small as possible, it was found later that the layered structures may provide a good platform to study topological insulators and related semimetals (Dirac and Weyl). The topological properties of iPCM rely on the crystalline structures, which satisfy or not the spatial inversion symmetry. If the iPCM has the symmetry, magnetic properties are not induced because the electronic band structures are degenerated. This is the cases for Kool, Petrov and inverted-Petrov phases. On the other hand, the only ferroelectric phase of iPCM breaks the symmetry, which lifts two different spin bands, resulting in magnetized. Interestingly, the magnetic properties attribute to p-electrons in Ge, Sb and Te, which are all non-magnetic elements.
Multiferroics is a keyword for future electronics as well as topological insulator. However, it is still difficult to induce large electrical and magnetic properties at room temperature. It relies on that electric dipole moments are usually related to p-electrons while the magnetic moments are to d-electrons. If a large magnetic moment is induced from p-electrons, a giant multiferroics would be possible.
A combination of a topological insulator and a ferroelectric insulator may open a new era to realize such the multifunctionality. Topological insulators, such as Bi₂Te₃ and Sb₂Te₃, usually satisfy both spatial inversion and time reversal symmetry. The topological surface bands are mainly made of the band inversion of 6p- or 5p-electrons of Te, Sb and Bi. On the other hand, GeTe is known as a ferroelectric material, which has a large spin-orbit coupling (SOC) compared with other oxide ferroelectric materials. Due to the large SOC, it shows a large Rashba-like spin split bands. It is noted here that the existence of an electric dipole moment breaks the spatial inversion symmetry.
If thin films of Sb₂Te₃ and GeTe are piled up alternatively, what happens on the band structure as the bulk film? Actually, both layers can share a lattice plane using (0001) and (111) through van der Waals force. As the bulk film, the spatial inversion symmetry is broken. Therefore, plural Dirac cones appear apart from the Γ-point in the k-space, resulting in a Weyl semimetal. Weyl semimetals are magnetic sensitive because spin bands are lifted from the band degeneration.
In the presentation, we show several experimental results of the Weyl semimetal from superlattices consisted of GeTe and Sb₂Te₃ sublayers at room temperature.

References:
1. J. Tominaga, P. Fons, A.V. Kolobov,T. SHima, T. C. Chong, R. Zhao, H. K. Lee, L. Shi, Jpn. J. Appl. Phys. 2008, 47, 5763.
2. E. Simpson, P. Fons, A. V. kolobov, T. Fukaya, M. Krbar, T. Yagi, J. Tominaga, Nat. Nano. 2011, 6, 501
3. H. Zhang, C.-X. Liu, X.-L. Qi, X. Dai, Z. Fang, S.-C. Zhang, Topological insulators in Bi₂Se₃, Bi₂Te₃ and Sb₂Te₃ with a single Dirac cone on the surface, Nat. Phys. 2009, 5, 438.
4. D. D. Sante, P. Barone, R. Bertacco, S. Picozzi, Electric control of the giant Rashba effect in bulk GeTe, Adv. Mater. 2013, 25, 509.
5. J. Tominaga, A. V. Kolobov, P. Fons, T. Nakano, S. Murakami, Ferroelectric order control of the Dirac-semimetal phase in GeTe- Sb₂Te₃ superlattices, Adv. Mater. Interfaces 2013, 1, 1300027.
6. J. Tominaga, Y. Saito, K. Mitrofanov, N. Inoue, P. Fons, A. V. Kolobov, H. Nakamura and N. Miyata, A magnetoresistance induced by a nonzero Berry Phase GeTe/Sb₂Te₃ chalcogenide superlattices, Adv. Funct. Mater. 2017, 1702243.

Epitaxially grown GeSbTe alloys in the metastable rock-salt phase and the stable trigonal phase are transferred in ultrahigh vacuum to employ angularly resolved photoelectron spectroscopy (ARPES) and scanning tunneling spectroscopy (STS). Firstly, we map the band structure of Ge₂Sb₂Te₅ [1] within the whole Brillouin zone, which exhibits valence band maxima away from the G point and well below the Fermi level. Hence, the Fermi level is pinned in the disorder broadened part of the bands constituting a so-called pseudo-Fermi-surface, where each enclosed state contributes only partially to the conductivity. The resulting hole density of this pseudo Fermi-surface agrees well with results from Hall measurements of the same sample type.
Moreover, via two-photon ARPES, we found an in-gap surface state with largely linear dispersion and a circular dichroism being mirror symmetric in k-space as expected for a topological surface state. Estimates, however show that the intrinsic bulk carrier density overwhelms the charge carrier density from this surface state significantly, such that the surface state is likely not relevant for the transport properties of metastable Ge₂Sb₂Te₅.
Using STS, we established methods to probe the subsurface GeSb layer, which includes vacancies and is located below the covering Te layer [2]. It turns out that the subsurface layer is relatively ordered such that individual defects can be spotted. Finally, we contact the GeSbTe layers with the STM tip and are able to switch the resulting circuit between states of different resistivity via applying voltages of about 1V.
[1] Kellner et al., Comm. Phys. 1, 5 (2018).
[2] Kellner et al., Phys Rev. B 96, 245408 (2017).

The ability to reprogram the interaction of visible light with nanostructures has a range of applications from bio and chemical detection to computation. However, many of the existing tuning mechanisms are slow and volatile. In contrast, phase change materials (PCMs) offer non-volatile programmability on sub-nanosecond time scales. Unfortunately the common phase change materials, which were originally developed for data storage, strongly absorb visible light and tend to exhibit a relatively small change to the real part of the refractive index, Re(n), in the visible spectrum. A large ΔRe(n) is required to tune the resonant frequency of photonics structures. Therefore, new phase change materials that exhibit a large ΔRe(n) must be designed specifically to transmit visible light.

Antimony sulphide (Sb₂S₃) is a somewhat misunderstood PCM. It exhibits a large ΔRe(n) between its amorphous and crystalline phases[1] and transmits visible light. It was originally explored in the 1990s for data storage applications but was soon forgotten because it seemed difficult to amorphise using laser pules[2]. At the time PCMs were being developed for rewritable compact discs (CDs) and digital versatile discs (DVDs), which operated at wavelengths of 780 nm and 658 nm respectively. The large band gap of Sb₂S₃ prevented it from efficiently absorbing the laser energy, which is required to heat the material and induce structural phase transitions. However, recenlty we designed resonator structures that strongly enhance the absorption of Sb₂S₃ and in turn allow it to be efficiently heated and subsequently amorphised on a nanosecond time scale using laser pulses[1].

We have now applied Sb₂S₃ to a variety of different optical resonator structures, which allow control of both the intensity and phase of the transmitted light. In this talk, the properties of Sb₂S₃ will be presented, and the performance of our latest reprogrammable visible photonics devices will be discussed.


1. Weiling Dong, Hailong Liu, Jitendra K Behera, Li Lu, Ray J. H. Ng, Kandammathe Valiyaveedu Sreekanth, Xilin Zhou, Joel K.W. Yang, and Robert E. Simpson, Wide band gap phase change material tuned visible photonics, aXiv Preprint (2018), no. 1808.06459.

2. P Arun and AG Vedeshwar, Phase modification by instantaneous heat treatment of sb2s3 films and their potential for photothermal optical recording, J. Appl. Phys. 79 (1996), no. 8, 4029–4036.

Reconfiguration of silicon photonic integrated circuits relying on the weak, volatile thermo-optic or electro-optic effect of silicon usually suffers from large footprint and energy consumption. Here, integrating a phase-change material, Ge₂Sb₂Te₅ (GST) with silicon microring resonators, we demonstrate an energy-efficient, compact, non-volatile, reprogrammable platform. By adjusting the energy and number of free-space laser pulses applied to the GST, we characterize the strong broadband attenuation and optical phase modulation effects of the platform, and perform quasi-continuous tuning enabled by the thermo-optically-induced phase changes. As a result, a non-volatile optical switch with high extinction ratio as large as 33 dB is demonstrated. We also demonstrated new device design to demonstrate low loss directional coupler, even though the material itself has very low loss. Finally, using thermal heater based on silicon p-i-n junction, we demonstrated electrical actuation of the phase transition. Few cycles of switching were demonstrated.

In this talk, I shall cover recent developments in the field of evanescently couple phase change materials-based photonic devices. Specifically, the use of such devices to carry out non-von Neumann computing approaches, where procssing and memory can be co-located, but on integrated photonic chips will be the primary topic. Specific examples related to matrix-vector multiplications as well as scaling up these techniques towards larger matrices, carrying out electro-optical converstions and reducing power and wafer real estate requirements will also be discussed.

The number of broadband infrared (IR) materials that are commercially available is very limited. This in turn limits the design of current Electro-Optical / Infrared (EO/IR) sensors; making them bulky and expensive to manufacture. However, next generation EO/IR sensors, will require more compact, lighter and cost-effective systems for applications where size, weight, power and cost (SWaP-C) becomes a limitation. Due to this fact, the design of next gen EO/IR sensors will require novel multi-functional materials, with tailorable optical properties that can be engineered to serve complex optical functions.
In this work we highlight the development of broadband IR materials including chalcogenide glass-ceramics and chalcogenide phase change materials with tailorable optical properties. The local refractive index of the material is modified by controlling the local concentration of nanocrystals in the glass matrix using a laser; thus, facilitating the fabrication of complex gradient index (GRIN) components. Consequently, we exploit this additional degree of freedom, not available in traditional IR materials, to develop more compact optical designs to enable next generation EO/IR sensors for SWaP-C limited applications.
Here we demonstrate the imaging power of our broadband GRIN materials by designing, fabricating and characterizing a flat GRIN lens. Typically, 1-in diameter coupons are readily fabricated. The progress in manufacturing readiness level (MRL) will also be discussed.
Finally, the properties of novel broadband phase change materials will be highlighted. This material exhibits a large refractive index change, in the order of 1.5, across the IR spectrum. The design, fabrication and characterization of reconfigurable flat metasurfaces will also be presented.

Chalcogenide semiconductor alloys offer a uniquely functional and compositionally-controllable material base for nanophotonic, plasmonic and optical-frequency metamaterial applications. They variously present high- and low-index dielectric, low-epsilon and plasmonic properties at near-UV to near-IR wavelengths, coupled to a capacity for fast, non-volatile, electrically-/optically-induced switching between phase states with markedly different properties. We present recent developments in their application to non-volatile reconfiguration in photonic metamaterials, including: switchable ‘structural colors’ underpinned by a transition between characteristically dielectric and plasmonic states; and the first optically-switchable UV/HEV dielectric metamaterials, wherein the functional chalcogenide is hybridized with a transparent, high-index dielectric supporting the resonant mode and phase-switching, unusually, changes resonance quality but not spectral position.

A periodic array of sub-wavelength apertures in a thin metal film exhibits high transmission (transparency) for a narrow wavelength band, with the majority of light otherwise reflected, a phenomenon known as extraordinary optical transmission (EOT) [1]. In this work, such a hole-array is coated in a phase-change material (PCM), specifically Ge₂Sb₂Te₅ or Ge₃Sb₂Te₆, whose refractive index changes the wavelength at which incident light can couple into the metallic hole-array. Since the refractive index of the PCM can be altered by changing its phase-state, the PCM layer enables the EOT transmission peak to be dynamically tuned (shifted) in wavelength, so providing the functionality of a tunable band-pass filter [2]. Moreover, such filters can work in almost any wavelength range for which a suitably low-loss PCM is available. Here we show filtering operation in the 3 to 5 µm and 8 to 12 µm infrared wavebands. Since such filters are ultra-thin, solid-state and can be continuously tuned, they are suited to multispectral imaging and display applications. They also possess large field enhancements within the holes, opening up the possibility for use as chemical sensors with enhanced sensitivity. Many advantages of phase-change materials in general transfer to these filters, namely fast (ns) nonvolatile (years) switching, with excellent cycling endurance [3]. It is also possible to add or remove additional characteristics to these filters, such as polarization or angular sensitivity, by carefully selecting the shape of the holes used to pattern the metallic layer [4].

For multispectral imaging applications, the PCM band-pass filter is placed in the optical path between a scene and a high-speed camera. Multiple images are then taken as the center wavelength of the filter is scanned continuously over the waveband of interest by progressively crystallizing the PCM layer. The advantages of this method are the wide wavelength range applicability, and that spectral resolution can be improved using a higher speed camera or tuning the filter less rapidly (by lowering the temperature used for crystallization of the PCM layer).

Another possible application that we explore is the provision auto-referencing chemical sensors that utilize infrared absorption spectroscopy for analyte identification. Here the filter’s resonance wavelength is designed to match the primary absorption peak of the analyte. The high field enhancement within the filter’s hole-array causes strong absorption at what would normally be the high transmission wavelength, so yielding high measurement sensitivity. The filter can then be tuned to a region in which the analyte does not absorb, and a baseline taken to allow for auto-referencing.

Finally, we consider a third potential application, that of infrared displays. Here, by layering one filter on top of another we can control both the transmission wavelength and the transmitted intensity. Such a configuration could thus be combined with a broad-band infrared source to produce sub-wavelength thick, kHz refresh rate, infrared displays capable of producing almost any arbitrary test image (e.g. for sensor calibration, target identification etc.).

References
[1] Ebbesen, T. W. et al. Nature 86, 1114–7 (1998).
[2] Rudé, M. et al. Adv. Opt. Mater. 4, 1060–1066 (2016).
[3] Kim, I. S. et al., 2010 Symposium on VLSI Technology; 203–204 (2010).
[4] Lin, L. and Roberts, A. Appl. Phys. Lett. 61109, 1–4 (2017).

Ge-Sb-Te (GST) phase change materials exhibit a structural phase transition between amorphous and two crystalline phases. With increase in temperature, amorphous GST transforms into cubic and hexagonal crystals at around 150 °C and 350 °C, respectively. During this phase change process, not only optical reflectivity, but also carrier density and mobility significantly enhance. Based on these properties, various kinds of new optical and electrical applications, such as optical memory/logic, display, and RF switch, have been proposed. However, optical properties in the terahertz (THz) frequency region have not been fully unveiled, although GST is expected to be utilized for THz optoelectronic device applications [1]. In this study, we employed a transmission-type THz time- domain spectroscopy (THz-TDS) and evaluated the optical properties of GST thin films in the THz frequency range.
The samples were Ge₂Sb₂Te₅ alloy films (100 nm) deposited on Al₂O₃ substrates. ZnS-SiO₂ protection layers (20 nm) were then deposited. The GST film was grown at room temperature and then annealed at 150 ~ 400 °C for an hour to obtain crystalline GST films. It was found that the absorption coefficient increases over the entire frequency range of the measurement (0.2 ~ 2.5 THz) with increase in the annealing temperature. The obtained flat transmittance curves are a characteristic of Drude absorption and indicate that the free carriers absorption governs THz properties. The observed trend can be qualitatively explained by the change in the carrier density and mobility. We found that the phase shift introduced by annealing is very small and thus GST alloy can be utilized for a programmable THz amplitude modulator as well as other THz devices by exploiting the temperature-dependent electrical property change.

[1] K. Makino et al., ACS Appl. Mat. & Interfaces 8, 32408 (2016).

We present the highly ordered array of vanadium dioxide (VO₂) nanodots using nanoporous templates fabricated by block copolymers-based lithography. Thin layers of poly(vinyl alcohol) (PVA) and polystyrene-block-poly(2-vinyl pyridine) (S2VP) are subsequently prepared on a silicon (Si), glass, and sapphire (Al₂O₃) substrate by spin-coating method. Then, the S2VP thin film on the PVA layer is solvent-annealed in toluene vapor for inducing its self-assembly, so that hexagonally packed poly(2-vinyl pyridine) (P2VP) cylinders oriented vertically to the surface are achieved. After the S2VP thin film is immersed in ethanol which is a selective solvent of P2VP blocks, the resulting film shows hexagonally packed nanoporous structures on its surface. Subsequently, the layer of VO₂ is uniformly deposited on nanoporous templates using vanadium (V) target with magnetron sputtering system. As the underlying PVA template as a sacrificial layer is removed by washing with water, only hexagonal array of VO₂ dots is remained on the surface. Since VO₂ is a phase-change material driven by temperature, electrical properties of these VO₂ dots are characterized by varying temperatures.

Much attention has been paid to next-generation memory devices capable of low-power consumption and high-density integration. We have recently reported a novel nonvolatile memory device (Voltage switching Random Access Memory: VolRAM) based on the change in the standard electrode potential.¹ The advantage of this device is the low-power consumption; however, the relation between the applied pulse voltages for switching and power consumption remains unclear. Accordingly, we studied memory operations using pulsed voltages, regarding power consumption.
The device consists of a stack of Li, Li₃PO₄ solid electrolyte, and Ni thin films on glass substrate. All the processes including fabrication and evaluation were performed without exposing samples to air using our custom-made system.² We first fabricated VolRAM devices with a stacked structure of Li(1 µm)/Li₃PO₄(1 µm)/Ni(100 nm) on glass substrates, with a typical diameter of 0.5 mm. For VolRAM operation, we applied rectangular wave pulsed voltages (maximum of 4.5 V vs. Li) between Ni-Li electrodes as the initial state with an open-circuit voltage of 0.95 V (Low-voltage state). Both of the pulse-width and waiting time between each pulse were set to 100 ms. The amount of Li migration during voltage application was calculated by the time integration of current observed in the circuit. After applying the pulsed voltages, the time dependence of the open-circuit voltage was measured.
The open-circuit voltage increased as the number of application pulses increased. The voltage states of VolRAM successfully switched from low-voltage state to intermediate- and high-voltage state by the applications with 2 and 20 pulses, respectively. Note that the amount of Li ions migrating in the two pulses was 26.1 pC/µm². If we assume this device to be the size of 20 nm × 20 nm, the power required for switching is 5 × 10^-14 J. This value is very small compared with the typical energy required by a DRAM device: 10^-13 J.³ Thus, we succeeded in the operation of VolRAM by applying pulsed voltages with very low power consumption.
[Reference]:
1: Sugiyama, Hitosugi et al., APL Mater., 5, 046105 (2017).
2: Haruta, Hitosugi et al., Solid State Ionics, 285, 118 (2016).
3: T.Vogelsang, 43rd Annual IEEE/ACM International Symposium on Microarchitecture, 43, 363 (2010).

The switching behavior of fast ion conducting (FIC) AgI-Ag₂O-MoO₃ glasses has been revisited with a quest for a better understanding of the structure-mechanism relationship and role of the electrode type. For that, we have thoroughly scrutinized the switching behavior for bulk AgI-Ag₂O-MoO₃ glasses over a wide range of compositions within the glass forming region, proportioning a balance between glass matrix former (MoO₃) and glass matrix expander (AgI). Samples with equal thickness of 0.2 mm and inert bronze electrode have been used for this purpose. The ion conduction mechanism of FIC glasses is fundamentally based on electric field driven thermally activated ion hopping model [1] and mixed iodine-oxygen coordination [2]. These two factors are basically structural in nature. A DSC study shows a monotonic increase in glass transition temperature (T_g) with MoO₃ concentration but the stability (T_c - T_g) profile, where T_c is the crystallization temperature exhibit a maximum plateau region around the central composition, (AgI)₅₀-(Ag₂O)₂₅-(MoO₃)₂₅. This result gives a very brief perception about the switching performance. The nature of the stability profile has been understood as a consequence of (i) AgI crystallization in the AgI enriched sample (ii) enhancement of network connectivity due to MoO₃ concentration increment. The switching study has been conducted with a triangular DC pulse with 1 mA maximum current with a step of 0.01667 mA (approx.) and a period of 0.125 s for each step. Unlike earlier studied work on AgI-Ag₂O-MoO₃ metal (electrode)-Insulator (electrolyte)-metal (electrode) (MIM) structured thin film samples [3], this present study exhibit a permanent latching into low resistance OFF state. No erase/RESET mechanism can alter the state. This happens because the switching mechanism suggests [4] formation of a metallic filament, which is in the present case made of silver. Only silver ions (Ag⁺) within the electrolyte take part in the conduction process. Due to the applied electric field, these cations move in a hopping manner that has been explained in the lights of Mott-Gurney model [1]. When in contact with the inert electrode, cathodic deposition takes place, forming a metallic silver channel. This phenomenon is analogous to a short between two electrodes, and thus neither a high impulse current nor a shift in electrode polarity can alter the state from OFF to ON (high resistance).

This study shows how and why a change in electrode type can modify the switching behavior. Furthermore, how the switching behavior can be improved, has been discussed with the help of seemingly scattered nature of threshold voltage (V_th)- MoO₃ concentration profile and Mott-Gurney Model. Besides, bubble formation near the anode, which has been observed in scanning electron microscopic (SEM) image, is suggestive to oxidation reaction that incentivizes us to rethink earlier speculation [5]. Overall, the focus of this work is to understand some crucial issues regarding the switching behavior of this specific material to enable us to exploit the material for further development in non-volatile memory technology.

References

[1] R. Waser, et al. Adv. Mater. 2632-2663 (2009) 21
[2] J. Swenson and S. Adams. Phys. Rev. B. 024204 (2001) 64
[3] X. B. Yan et al. J. Appl. Phys. 054501 (2009) 106
[4] Solid State Electrochemistry II: Electrode, Interfaces and Ceramic Membranes. Edited by Vladislav V. Kharton. ISBN: 978-3-527-32638-9. (2011)
[5] B. Vaidhyanathan, S. Asokan, and K. J. Rao. Bull. Mater. Sci. 301-307 (1995) 18

In this work, three phase change memory (PCM) devices using layered chalcogenide materials, GeTe/SnTe, Ge₂Se₃/SnTe, Ge₂Se₃/SnSe, and one single material layer device comprised of Ge₂Sb₂Te₅ (GST) were fabricated and electrically characterized. The GeTe and Ge₂Se₃ bilayer devices have previously demonstrated lower threshold voltages and improved cycling reliability compared to the GST-based devices [1]. Part of this improvement was attributed to better mechanical electrode contact as well as electrical band alignment between the top electrode and the device material layers. To further explore the mechanisms responsible for the observed functional improvements, in this work the conduction mechanisms of all four device types were investigated using quasi-static DC I-V curve measurements over the high resistance state to the low resistance state, in the temperature range of 5 K to 340 K. Mechanisms investigated were Schottky emission, Poole-Frenkel emission, Fowler-Nordheim mechanism, direct-tunneling, space-charge limited conduction (SCLC), hopping, and extended states conduction.
Acknowledgments
Antonio S Oblea participated in data collection. Funding was in part by NASA grant NNX07AT60A.
Reference
[1] K. A. Campbell, C. M. Anderson, “Phase-change memory devices with stacked Ge-chalcogenide/Sn-chalcogenide layers,” Microelectronics Journal 38 (2007) 52-59.

Additive technology, popularly known as 3D printing, has been of great interest in recent years due to its several advantages over conventional fabrication technologies (low cost, high throughput, high conformity etc.). In the field of electronics, this technology has been specifically successful in printing metal electrodes and wires on flexible substrates. Inkjet, Aerojet or syringe dispensing systems are the different varieties of this technology where each of them requires ink with specific parameters (particle size, viscosity, surface tension etc.). Post-printing processing of the thin films involves sintering/curing which could need high temperature and must be optimized for not only different printing methods but also for different inks and specific applications.
As 3D printing can create arbitrary structures, it will be invaluable if the technology can be applied to print optical waveguides, gratings or oscillators. To take a step towards this, we report a step-by-step process for inkjet printing of chalcogenide glass starting from glass synthesis. Chalcogenide glasses (ChGs) are phase change materials that are widely used in optical and electrical memory because of their high refractive index, transparency in infrared (IR) region, high resistivity contrast between two phases and thermally driven amorphous-to-crystalline phase change. As an amorphous material, they have a remarkable tolerance to impurities, compositional variance and radiation-induced damage. Their phase transition to a crystalline condition caused high interest due to the application in the electronic or optical recording. While this requires development and synthesis of low-temperature switching mechanisms, recent advancements in ChGs application as temperature sensors in nuclear facilities demand glasses with high crystallization temperature (T_c) and high refractive index (n). High coordination number and relatively strong chemical bonds of Ge containing binary ChGs make them promising candidates in that aspect. The printing process starts with glass synthesis and then nanoparticles (diameter <= 100 nm) are obtained by an optimized wet milling of the bulk glass followed by ultrasonication. Cyclohexanone is used to suspend the nanoparticles and to avoid agglomeration, ethyl cellulose is added as a surfactant. Once the particle size and viscosity (2-30cP for Dimatix) are characterized by a DLS system and a Rheometer respectively, the ink is used in a Dimatix inkjet printer to obtain printed thin films. Printing requires optimization of the ink parameters, as well as the piezoelectric nozzle controlling waveform. Development of these specific parameters and the sintering process optimization for obtaining of homogeneous thin films are described in details and discussed based on the ChG glasses composition. Refractive indices and extinction coefficients of the thin films were measured as the films were produced and after heating at various target temperatures up to their crystallization temperature directly in a hot stage in an ellipsometer using light of 270 nm to 1650 nm. The surface morphology of the films was studied using scanning electron microscopy (SEM) and atomic force microscopy (AFM).

Metal-insulator transition (MIT) in strongly correlated materials has enormous potential with scientific and technological impacts in future oxide nanoelectronic devices. While photo-driven MIT can provide opportunities to extend the novel functionality of strongly correlated materials, there have rarely been reports on it. Here we report MIT provoked by visible-near infrared (Vis-NIR) light in Ag-decorated VO₂ nanorod arrays (NRs) due to localized surface plasmon resonance (LSPR). Under Vis-NIR illumination, strong electric fields, generated at the interface between Ag nanoparticles and VO₂ NRs by LSPR, affect the electronic state of VO₂ and provoke MIT and structural phase transition in VO₂ NRs. Due to LSPR-induced MIT in VO₂ under Vis-NIR illumination, the Ag-decorated VO₂ NRs exhibit MIT-driven broad spectral response with high and fast response. Our study will open up the new strategy to trigger MIT, a new functionalization in strongly correlated materials and serves as a valuable proof-of-concept for next generation optoelectronic devices with fast response, low power consumption, and high performance.

Phase change materials (PCMs) are currently being explored for use in a number of photonics applications due to their interesting characteristics. These materials have the ability to switch to multiple different states in a non-volatile fashion. These states can have drastically different optical characteristics. By controlling the state of the material, the amplitude and phase of light passing through or reflected off the material can be controlled. This can lead to the creation of many different photonic devices including high-speed light modulators and optical memories. We characterize the temporal dynamics and quantity of crystallization of Ge2Sb2Te5 (GST) utilizing a pump-probe setup. The pulses to facilitate the switching are created via a CW 1550 nm NP Photonics laser, an arbitrary waveform generator (AWG), and an electro-optic modulator (EOM). The pulses are then amplified by an erbium doped fiber amplifier (EDFA). Post EDFA, the probe laser (a 1580 nm CW Agilent laser amplified by a second EDFA) is coupled into the same fiber using an 80/20 fiber splitter. The probe power needs to be low enough that there is not extra heating of the sample, which would affect the phase change characteristics, but high enough to be detectable after reflection. The resulting mixed signal is then passed through a free-space EOM (used to pick off an arbitrary number of pulses) and coaligned with a 780 nm laser that is used for visually positioning the pump and probe lasers on the sample. The light then is sent into a microscope where it is incident on the GST on silicon sample. The reflected light is coupled back into a fiber and passes through two wavelength division multiplexers (WDMs) in order to separate out the probe signal from the pump before being detected by a 3.5 GHz photodetector and recorded on a digital phosphor oscilloscope with a resolution of 2 ns. As the phase is changed via the pulses, the probe laser sees a change in the reflectivity. The amount of change in the material can be determined by the reflectivity measurement from the probe laser and the temporal dynamics can be gleaned from the resulting graph. In our results we see multiple levels of reflectivity indicating partial change of the GST. After a number of pulses (dictated by the power of each pulse), the reflectivity levels off indicating that the GST is fully changed up to the penetration depth of the laser. From the shape of the graph the cooling times as well as information about the mechanisms of the change can be inferred.

The amorphous phase in the reset state of GST-based phase-change memory (PCM) cells shows a resistance increase in time according to power law [1] with an exponent dependent on the size of the programmable region.
This large increase in resistivity with time (resistance drift) during the ageing of the amorphous phase has up till now limited the multi-level storage capability, i.e. hindered the development of ultrahigh multi-level storage devices.
In explanations of the phenomenon of resistivity increase in the reset state of these PCMs published to date, a significant role has been attributed to the creation and growing of subcritical crystalline nuclei inside the amorphous area.
The results of analysis of the impact of the kinetics of the increase of the population and evolution of subcritical nuclei on the phenomenon of resistivity increase of the amorphous phase in reset state of PCMs are presented here. They are based on previously published research results for compound in the GeTe-Sb₂Te₃ tie line (e.g. [2], [3] )
.
[1] Li J, Luan B and Lam C 2012 IEEE Int. Reliab. Phys. Symp.Proc. p 1; [2] P. Noé et all , 2018, Semicond. Sci. Technol., 33,013002 (32pp); [3] J. Im et all 2011 Curr. Appl. Phys. 11, e82–4

Controlling light-matter interactions enables a wide range of applications from optical sensing to imaging and communications. However, for a given nanostructure, the optical response from the light-matter interaction is usually fixed after the device is fabricated. The optical response can be modulated by changing the nanostructure mechanically or by using materials with controllable refractive index. Phase-change materials exhibit a huge property contrast between their structural phases. And the phase transition can be typically invoked in nanoseconds. So far, the most commonly used phase change material for tuneable photonics research is Ge2Sb2Te5. The problem with Ge2Sb2Te5 is that its absorption coefficient in the visible spectrum is very high whilst the change in the real component of the refractive index in the visible spectrum is relatively small. This limits the usefulness of Ge2Sb2Te5 in the visible spectrum. In contrast, Sb2S3, a chalcogenide that has not been studied widely, is a good alternative for visible photonic devices due to its fast switching speed, large optical band gap and large refractive index change. We have shown that by using Sb2S3 in a simple optical resonator, the resonant frequency can be easily tuned by inducing a phase transition in Sb2S3 [1]. This phase transition is induced by optical or electrical heating. By changing the thickness and structural state of Sb2S3, the resonant wavelength of the structure can be tuned to cover the whole visible spectrum. Indeed, the maximum resonance shift was 110 nm [1]. Note that this simple structure is easily manufactured by industrially scalable techniques. We conclude that Sb2S3-based tuneable photonics is promising for a broad range of applications, such as reconfigurable waveguide systems, spectrometer, and display technologies.
[1] W. Dong, H. Liu, J. K. Behera, L. Lu, R. J. H. Ng, K. V. Sreekanth, X. Zhou, J. K. W. Yang, R. E. Simpson, Adv. Funct. Mater. 0, 1806181.

Acknowledgements
This research was performed under the auspices of the SUTD-MIT International Design Center (IDC). The research project was funded by the Samsung GRO, the A*STAR Singapore-China joint research program Grant No. 1420200046, and the SUTD Digital Manufacturing and Design Centre (DManD) Grant No. RGDM 1530302. The authors are grateful for fruitful discussions with Seokho Yun.

The emergence of data-driven computing has prompted a search for systems based on new computational elements. One novel paradigm is the use of spikes, or neuron-like action potentials, to communicate and compute. However, implementing such a neuromorphic system in physical hardware requires compact devices that can produce controllable, low-energy, and fast spiking.
One approach is to exploit instabilities in the transient dynamics of VO₂, which produce periodic spiking when used in a relaxation oscillator, corresponding to periodic heating and cooling across its Mott metal-insulator transition. Speeding up these devices requires aggressive shrinking to minimize both the electrical and thermal time constants of these oscillations. To effectively scale such VO₂ devices to sub-10 nm widths, in previous work we used metallic carbon nanotubes (CNTs) with ~1 nm diameter as nanoscale heaters in physical contact with an electrically parallel thin film of VO₂ [1].
In this work we demonstrate that a single CNT-VO₂ device forms a Pearson-Anson relaxation oscillator when driven by a DC source, generating neuron-like periodic spiking. The highly localized Joule heating of the CNT acts to trigger and confine the insulator-metal transition in the VO₂ to nanoscale dimensions, greatly reducing the effective thermal mass of the device. This reduction of thermal time constant allows the device to oscillate very rapidly across its Mott transition. Compared to a VO₂-only device, the CNT-VO₂ devices showed a reduction in quasi-static switching voltage accompanied by the following dramatic changes in its dynamical behavior: (a) an increase in the spiking frequency by over 3 orders of magnitude, (b) a decrease in the spike’s transient duration by 3-4 orders of magnitude, and (c) a decrease in pulse energy by 2 orders of magnitude. We also constructed a compact model that could quantitatively reproduce these observations, providing insights into the thermal oscillation effects induced by the CNT.
We further characterized the tunability and scalability of our CNT-VO₂ devices. We found that the spiking frequency of a single device can be tuned by nearly an order of magnitude by adjusting the DC bias conditions. Our results also showed that the spike width and energy decrease with length, with our shortest 300 nm devices showing sub-20 ns spike widths.
In summary, we demonstrated that the addition of a CNT nanoscale heater results in a significant improvement in the dynamical behavior of VO₂-based oscillators. This provides an accessible path to scaling electronic devices by thermally engineering and localizing the dynamics of an otherwise bulk transition mechanism.
[1] S. Bohaichuk, M. Muñoz Rojo, G. Pitner, C. McClellan, F. Lian, J. Li, J. Jeong, M. Samant, S. S. P. Parkin, H.-S. P. Wong, E. Pop, Device Research Conference, IEEE, June 2018. DOI: 10.1109/DRC.2018.8442223

Reducing the phase transition temperature of vanadium dioxide (VO2) to near-room temperature is necessary many thermochromic applications including space radiators. In this work, we demonstrate fabrication of thin films of VO2 doped with fluorine and tungesten via an aqueous sol-gel method. These films are incorporated into multilayer dielectric stacks that optimize the optical properties of two decades of the electromagnetic spectrum (0.4 - 40 microns) to maintain high solar reflectivity and modulate their emissivity in response to variable heat loads. By doping the films, the phase transition temperature of VO2 can be reduced from 68 C to less than 30 C, making this technology compatible with both manned and unmanned spacecraft.

Previous investigations on doping of VO₂ to enable tuning of transformation temperature, hysteresis width, and magnitude of the metal-insulator transition (MIT) in VO₂ have primarily focused on substitutional dopants. In a novel post-synthetic diffusive annealing technique, Boron has been introduced as an interstitial dopant which both depresses the phase transition temperature (10 ^oC/at % B) and introduces an increase of the heating transition temperature as a function of time and temperature below the transition temperature i.e. a thermal history dependence of hysteresis width. Here, the origins of this unique kinetic hysteresis are proposed to stem from B dopant relaxation within the host lattice following the crystallographic phase transition (low-temperature M1 to high-temperature R phase) and will be discussed in the context of structural characterization (near edge x-ray absorption) and density functional theory calculations. Furthermore, in order to probe the effects of B-doping on the resistivity of VO₂, two-terminal devices were photo-lithographically fabricated from single crystal B-doped VO₂ particles with a range of Boron concentrations. Electrical characterization demonstrated that introduction of Boron lowers the resistivity of the low-temperature M1 phase thereby reducing the magnitude of the change in resistivity across the MIT, while introducing a unique kinetic dependence of resistivity. Finally, current pulses will be used to electrically trigger the phase transition to demonstrate viable device behavior. As well as presenting a novel system in which to explore the effect of interstitial dopants on kinetic hysteresis, B-doped VO₂ devices demonstrating varied, kinetically accessible resistance states could be of use in neuromorphic computing applications.

Control over the concurrent occurrence of structural (monoclinic to tetragonal) and electrical (insulator to the conductor) transitions presents a formidable challenge for VO2-based thin film devices. Speed, lifetime and reliability of these devices can be significantly improved by utilizing solely electrical transition while eliminating structural transition. We design a novel strain-stabilized isostructural VO2 epitaxial thin-film system where the electrical transition occurs at ~77 oC without any observable structural transition. A thin-film heterostructures with a completely relaxed NiO buffer layer has been synthesized allowing a complete control over strains in VO2 films. We discover that these isostructural (10±1 nm) epitaxial VO2 films exhibit structural pinning of the monoclinic (insulator and metallic) phase due to bandgap changes. This structure is robust and stabilized throughout the whole temperature regime of 25-200 oC as the insulator-to-metal transition occurs. On the contrary, the completely relaxed thick (250±1 nm) epitaxial VO2 films undergo usual monoclinic (insulator) to tetragonal (metal) transition. The strain trapping in VO2 thin films occurs below a critical thickness by arresting the formation of interface misfit dislocations in accordance with domain matching epitaxy paradigm. Importantly, strain-trapped films with the stabilized isostructural atomic arrangement, consistently exhibit the conventional insulator to metal transition behavior with temperature, which is attributed to changes in electronic orbital occupancy and electron-electron correlations strongly fostered by the captive strain. Using density functional theory, we calculate that the strain in monoclinic structure reduces the difference between long and short V-V bond-lengths () in the monoclinic structure which leads to a systematic decrease in the electronic bandgap of VO2. This decrease in bandgap is additionally attributed to a ferromagnetic ordering in monoclinic phase to facilitate a Mott insulator without going through the structural transition.

In today’s data-centric world, where some of the most useful computing tasks are to extract meaningful information from massive amount of unstructured data, neuromorphic computing can provide low-energy high throughput computing. The challenge in data-centric computing with the conventional computing architecture lies in the energy and latency bottleneck of off-chip memory access (i.e. “memory wall”) which do not scale down with the scaling of the technology node.
Deep neural network (DNN) is a class of artificial neural networks (ANNs) that benefits from both the availability of big data (large amount of multi-media data for model training) and the continual performance improvement of semiconductor technologies in the past decade. However, the memory hiararchy of today’s computing architectures is not specifically designed to leverage the predictable dataflow and potential data reuse of DNN processing, resulting in longer latency and insufficient energy-efficiency for memory access. To reduce these expensive memory access, hardware accelerators for DNNs are designed to employ more fine-grained local memory hiararchy and more specialized dataflow design, which improves the energy efficiency and throughput while maintaining DNN’s inference accuracy. However, the "memory bottleneck" in modern DNNs may not be fully addressed by the acceleration architectures alone. Emerging memory technologies have the potential to play an important and unique role. As these technologies can potentially offer up to tera-bytes of on-chip data storage with a wide range of energy-delay optimization opportunities, they may complement SRAM and DRAM for more efficient DNN inference acceleration. A possible application of the emerging non-volatile memory (NVM) devices is to serve as in-memory computing element where multi-level resistance response of an NVM can store the analog synaptic weights of a DNN on-chip. In another in-memory computing scheme, a crossbar array of non-volatile memory devices can perform the multiply-and-accumulate (MAC) operation at a lower energy cost when the input vector is encoded as an analog voltage and the weight matrix is encoded as analog resistance (conductance) values stored in the memory devices. The ability of the NVMs like RRAM, PCM, CBRAM to change its resistance values gradually as a function of the applied voltage pulse across its electrode is the key to performing analog in-memory MAC operation.
This paper provides an overview of the current state-of-the-art non-volatile memory devices used for neuromorphic hardwares in applications ranging from biology based learning models to conventional machine learning algorithms solved using neural networks. Furthermore, a more focused overview of the device-level trade-offs required for hardware acceleration of neural network architectures using analog in-memory MAC operation is presented. In general, larger conductance range, more intermediate states, and higher resistance are desirable for both inference and traning. For inference, an ideal device should also have linear I-V relationship and long retention time. For training, symmetric and linear pulse response, small device-to-device and cycle-to-cyle variation, and good endurance are crucial.
Our review reveals that controlling the oxygen ion movement during pulsed switching in RRAM can be a promising way to achieve the aforementioned performance goals. Placing an oxygen ion barrier to make a bilayer RRAM and confinement of the generated heat during switching have shown significant improvement in analog switching. Better thermal management in RRAM can also provide filament stability that could improve reliability like retention and endurance. If the ideal device can be achieved, the MAC array using NVMs can provide ultra-low energy, high throughput computing without compromising bit precision that is currently missing in the neural network accelerator landscape.

The conventional computing systems based on the von Neumann architecture have reached limits in terms of computational time and power dissipation due to increasing computational complexity.¹ To circumvent these issues, new computational architectures are explored, among which neuromorphic computing based on emulating the human brain stands out. It is known that the human brain supercedes a supercomputer by 6-9 orders of magnitude in power dissipation. The superior features of the brain, such as ultra-high density, low-energy consumption, parallelism, robustness, plasticity, and fault-tolerant operation need to be emulated by computing systems for perception and learning.¹ These qualities are enabled by the brain’s synapses which form a highly complex and efficient interconnection between neurons in the brain. Therefore, a nano-electronic device which emulates the synaptic properties is a crucial building block for brain-inspired computational systems.
It is observed that the programming current in a memristor is decreased by using 2D materials^2-3 as the switching medium, and by using graphene electrode instead of conventional metals.^4,5 However, these ultra-low power memristors are realized by exfoliated 2D materials which makes large-scale production implausible.^2-3 In this work, we use MoS₂ as the switching medium to fabricate a memristive synapse. MoS₂ is grown on monolayer graphene forming a vertical heterojunction by the sulfurization of Mo film. Graphene (Gr) is used as the bottom electrode while Ni/Au contact on MoS₂ is the top electrode.
The MoS₂/Gr memristors exhibit forming-free non-volatile switching behaviour with a low programming current of 1 nA. In addition, these devices exhibit sub-nW reset power and low energy consumption of ~2 pJ/spiking event which makes them highly energy-efficient. Synaptic characteristics such as multiple conductance states between 1 nA and 1 μA with a minimum reset current of 16 pA is also observed. Furthermore, the MoS₂/Gr memristors exhibit excellent data retention characteristic of 10⁴ s at multiple conductance states viz., 1 nA and 1 μA. Also, MoS₂/Gr memristors exhibit biological synaptic characteristics like plasticity, multi-state conductance tuning, short and long-term potentiation, long term depression and spike timing dependent plasticity rule. Additionally, sustained switching at 1 nA and 100 nA programming currents is observed in these devices for 100 DC cycles indicating their robustness. Filamentary switching with low programming reset power observed in these devices unravel the prospects of developing nano-scale crossbar arrays of energy-efficient synaptic devices for neuromorphic applications.

1. Kuzum, D.; Jeyasingh, R. G.; Lee, B.; Wong, H.-S. P., Nanoelectronic programmable synapses based on phase change materials for brain-inspired computing. Nano Lett. 2011, 12 (5), 2179-2186.
2. Tian, H.; Zhao, L.; Wang, X.; Yeh, Y.-W.; Yao, N.; Rand, B. P.; Ren, T.-L., Extremely Low Operating Current Resistive Memory Based on Exfoliated 2D Perovskite Single Crystals for Neuromorphic Computing. ACS Nano 2017, 11 (12), 12247-12256.
3. Zhao, H.; Dong, Z.; Tian, H.; DiMarzi, D.; Han, M. G.; Zhang, L.; Yan, X.; Liu, F.; Shen, L.; Han, S. J., Atomically Thin Femtojoule Memristive Device. Adv. Mater. 2017, 29 (47).
4. Tian, H.; Chen, H.-Y.; Gao, B.; Yu, S.; Liang, J.; Yang, Y.; Xie, D.; Kang, J.; Ren, T.-L.; Zhang, Y., Monitoring oxygen movement by Raman spectroscopy of resistive random access memory with a graphene-inserted electrode. Nano Lett. 2013, 13 (2), 651-657.
5. Chakrabarti, B.; Roy, T.; Vogel, E. M., Nonlinear Switching With Ultralow Reset Power in Graphene-Insulator-Graphene Forming-Free Resistive Memories. IEEE Electron Device Letters 2014, 35 (7), 750-752.

Ferroelectrics as phase-transition materials have promising applications in both logic and memory technologies beyond CMOS. Neuromorphic architectures, such as neural networks and artificial brains [1][2], requires synergy among neurons and synapses to achieve cognition and classification functionalities. In spite of the fact that traditional CMOS technologies provide cumbersome topologies to achieve these functionalities, emerging devices, such as ferroelectrics field-effect-transistors (FEFETs), offer significant benefits in terms of power, performance, and area as physical platform for implementing neural networks due to their unique properties—hysteresis, non-volatility, multi-state, and etc. In the past a few years, FEFETs have demonstrated great potential in implementing artificial synaptic devices [3]-[5]; however, the demonstration of artificial neurons built on ferroelectric devices is still pending. In this talk, we will present experimental demonstration of ferroelectric spiking neuron based on a compact 1T-1FEFET structure and projected performance of spiking neural network based on ferroelectric neurons for unsupervised clustering on MINST dataset [6]. Ferroelectric spiking neurons utilize abrupt hysteretic transition feature of ferroelectrics such that there are unstable states in the current-voltage characteristics of FEFETs [7]. FEFET hysteresis can be dynamically tuned by bias conditions which allows for inhibition functionality. Artificial neurons based on other emerging devices, such as metal-insulator-transition (MIT) devices [8], suffer from a fundamental shortcoming--all these neurons are excitatory. In contrast, ferroelectric spiking neurons have built-in excitatory and inhibitory input connections, which are essential to enable high accuracy in unsupervised learning, increases sparsity in spiking, and efficient implementation of synaptic weights.
In summary, spiking neurons based on FEFETs provide compact and efficient implementation of artificial neurons and have great potential in developing spiking neural networks and other neuromorphic applications.

Reference:
[1] A. Merolla, et al. Science, vol. 345, no. 6197, pp. 668-673, 2014.
[2] C. Eliasmith, et al. Science, vol. 338, no. 6111, pp. 1202-1205, 2012.
[3] Y. Nishitani, et al. J. Appl. Phys., vol. 111, no. 12, pp. 124108, 2012.
[4] M. Jerry, et al. 2017 IEEE International Electron Devices Meeting, 2017.
[5] S. Oh, et al. IEEE Electron Device Lett. Vol. 38, no. 6, pp. 732-735, 2017.
[6] Z. Wang, et al. 2018 IEEE International Electron Devices Meeting, 2018.
[7] Z. Wang, et al. IEEE Electron Device. Lett., vol. 38, no. 11, pp. 1614–1617, 2017.
[8] M. Jerry, et al. Proc. 2017 Symp. on VLSI Tech., pp. T186-T187, 2017.

When executing neural algorithms, neuromorphic computers can overcome efficiency bottlenecks inherent to digital computers by parallel processing of synaptic weights in memory. Despite the possibility of improved efficiency, accurate and parallel programming is challenging to realize due to strict requirements for high write linearity and low write current. Redox transistors based upon lithium-intercalation oxides[1] or semiconducting polymers[2] combine linear programming and low write current in a single device. However, programming at the array level requires an additional selector device. Here, we demonstrate that the sharp ON threshold and high OFF resistance of a Ag-based selector[3] enables addressable programming of redox transistors and state retention. Combined, these two devices comprise an ionic floating-gate memory (IFG) similar to flash but with electrochemical mechanisms for charge injection and memory storage. With an IFG crossbar array we demonstrate accurate and parallel synaptic weight updates that are "blind" and therefore require no feedback mechanisms or multi-pulse schemes. Prototype IFG devices exhibit >1 MHz switching rates, >10⁹ write-read endurance and read currents of 5-10 nA that could enable scaling to crossbars with > 1000×1000 elements. Architectural analysis projects a hybrid IFG-CMOS accelerator to have an energy, latency, and area advantage of 476×, 16×, and 9.5×, respectively, compared to an optimized 8-bit digital accelerator.
[1] E. J. Fuller, et al., Advanced Materials, 2017
[2] Y. van de Burgt, et al., Nature Materials, 2017
[3] Midya, et al., Advanced Materials, 2017

We demonstrated gradual conductance update during the RESET process of Ti/HfO₂/TiN memristive device under different device conductance update schemes. These schemes include varying voltage pulse amplitudes and pulse widths. The reported devices exhibited ~50 dynamic ratio of conductance with switching speeds of 100ns and 400ns for SET and RESET processes respectively. 7 different voltage amplitudes were used to switch the devices from their highest conductance to lower conductance at 1000 cycles each. It was observed that the distribution of these 7 conductance states overlapped with one another, owing to the stochasticity nature of device switching dynamics. One of the most critical stochasticity aspects of memristive devices is the presence of random telegraph signal (RTS), in the form of either bimodal or multilevel current fluctuation during read operations. Despite its undesired properties resulting in conductance instability during read operation, bimodal RTS generated from the back and forth movement of an electron between one electrode to an intrinsic oxide trap has been proven useful in characterizing traps behaviour and location in oxide-based memristive devices. Based on classical theory, the electron capture and emission time of this specific RTS have opposite bias polarity dependence. In the capture of this RTS, the measurement was performed by varying the read voltage from ±0.02V to ±0.20V in steps of ±0.01V under different sampling rates, i.e. 8 ksps, 80 ksps, and 800 ksps for each of the write sequence. The measured signals were investigated by analyzing its time lag plot and bias dependence polarity of capture/emission time constants. It was observed that at higher read voltages, the probability of multilevel RTS occurrences increased (from ±0.15V to ±0.20V), hence the read voltage was optimized in the lower regime rather than the higher regime. The absolute value of RTS amplitude was found to increase as the read voltage increased, while its relative percentage to the mean current decreased. The bimodal signal with opposite polarity dependence is further analyzed to extract the defects location in the oxide structure. The estimated defects location before and after the writing process indicated the movement of the defects within the oxide region, resulting in the conductance change of the device. Different jump distances were extracted from the two conductance update schemes to provide an insight on its correlation with gradual conductance change in Ti/HfO₂/TiN memristive device.

Continuous enhancement of performance and processing powers of computing devices will soon reach a technological and physical limit and to overcome this, systems emulating the human brain are being developed. Presently, mathematical models known as artificial neural networks (ANNs) are designed to simulate the computational abilities of the brain. Due to large simulation times, the current digital ANNs cannot be efficiently scaled. By contrast, analog (ad opposed to digital) neuromorphic devices with dedicated and adaptable synapses are expected to rival the scale and efficiency of the brain. Phase change materials (PCMs) enables us to fabricate synapses through their memristive behavior. Memristors are two-terminal electrical component where the resistance is a function of the amount and direction of current. Here, we discuss memristive behavior of W/WO₃/Ag, Cu/CuWO₄/W, and Ti/HFO₂/Pt systems for neuromorphic architectures and their potential application as ANNs. In prior studies, memristive architectures for neuromorphic computing are connected through a crossbar array of neurons. With a limit on the number of neurons and regular connectivity, the networks lack sparsity and randomness. As a result, they have high wiring cost and poor functional connectivity. To address such limitations, we have fabricated a random array of core-shell memristive wires to form the connectivity matrix for neuromorphic hardware. In the systems listed above, the conductive core (W or Ti) serves as the bottom electrode with a memristive shell (WO₃ or HfO₂) that can be electroformed with a set of top electrodes (Ag or Pt) deposited on the surface. Prior to miniaturization, the core-shell wire networks were fabricated using 20 μm tungsten fibers as conductive cores with tungsten oxide as memristive shells. After imprinting of top silver electrodes and electroforming, IV measurements revealed memristive behavior with switching between resistive states (i.e., LRS-HRS) at ±1 V. Subsequently, nanowire networks based on the Ti/HFO2/Pt system were fabricated using electron beam lithography (EBL). Over 90 overlapping core (40 nm)-shell (5 nm) nanowires were written on a 200 μm × 200 μm write field. To complete the connectivity matrix, 64 nodes and individual vias connecting the resulting network to electrode pads were written and deposited. Network quantifying simulations revealed a small-world coefficient of 2.89, shortest path length of 3.61 and clustering coefficient of 0.057. When electroformed, this system exhibited switching between LRS and HRS at ±7 V. To investigate the role of electrode-memristor interfaces, copper tungstate films were deposited on copper substrates using chemical solution deposition. After sputtering top tungsten electrodes and electroforming, IV measurements revealed memristive behavior and exhibited switching between LRS and HRS at ±0.7 V. In this presentation, the fabrication process and the memristive characteristics (endurance, resistance retention, pulse measurements, etc.) of the above listed systems will be discussed in detail.

Inspired by neural computing, the pursuit of ultralow power neuromorphic architectures with highly distributed memory and parallel processing capability has recently gained more traction. However, emulation of biological signal processing via artificial neuromorphic architectures does not exploit the immense interplay between local activities and global neuromodulations observed in biological neural networks, and hence are unable to mimic complex biologically plausible adaptive functions like heterosynaptic plasticity and homeostasis. Here, we demonstrate emulation of complex neuronal behaviours like heterosynaptic plasticity, homeostasis, association, correlation and coincidence in a single neuristor via a novel dual-gated architecture. This multiple gating approach allows one gate to capture the effect of local activity correlations and the second gate to represent global neuromodulations, allowing additional modulations which augment their plasticity and enabling higher order temporal correlations at a unitary level. Moreover, the dual-gate operation extends the available dynamic range of synaptic conductance while maintaining symmetry in the weight-update operation, expanding the number of accessible memory states. Finally, operating neuristors in the sub-threshold regime enables synaptic weight changes with high gain, while maintaining ultralow power consumption of the order of femto-Joules.

Many approaches to induce the semiconductor-to-metal phase transition in two-dimensional (2D) crystals via strain involve global strain of the entire substrate as opposed to local strain of individual devices. In this work, we present a new approach to induce the transition where individual devices are addressed via field-effect modulation of an ionomer (i.e., single-ion conductor). Specifically, strain is induced in each device using a voltage applied to the top gate of an electric-double layer (EDL)-gated, field-effect transistor (FET). The mechanism to induce strain is the same mechanism responsible for actuation in ionic polymer metal composites (IPMCs), which mimic the behavior of muscle. The active layer in an IPMC is a single-ion conductor, which deforms under applied voltage due an electrostatic imbalance created by the formation of an EDL on one polymer/electrode interface but not the other. The imbalance occurs because only cations, but not anions, are free to diffuse in the electrolyte. Modeling indicates cation densities up to 5 x 10¹⁴ ions/cm² at the ionomer/electrode interface under 2V applied gate bias, which is theoretically predicted to induce several percent strain on the 2D crystal – enough to induce the phase transition. Experimentally, a custom-synthesized ionomer is used to electrostatically gate both graphene and MoTe₂ FETs. Compared with backgating through SiO2, transfer characteristics on both EDL-gated graphene and MoTe₂ FETs reveal a similar enhancement of the n-branch using both the ionomer and a dual-ion conductor (i.e., one with mobile cations and anions). The ionomer decreases the subthreshold swing in the n-branch of the MoTe₂ FET from 5000 to 250 mV/dec and increases the current density and on/off ratio by two orders of magnitude. However, as expected, the ionomer quenches the p-branches in both the graphene and the MoTe₂ FETs because the anions are immobilized and therefore cannot form an EDL by inducing holes in the channel. This is the first demonstration of an ionomer-gated 2D FET and it lays the groundwork for demonstrating the phase transition to create a steep-subthreshold swing device. EDL response speed will also be discussed, with modeling and experimental data supporting nanosecond response times by aggressive scaling of the device geometry.

This work is supported by the NSF-DMR under grant #1607935.

Phase change memory is a resistive non-volatile memory technology that stores information in the crystalline (conductive) or amorphous (resistive) phase of a material such as the chalcogenide Ge₂Sb₂Te₅. Phase change memory can be implemented as a crossbar array, allowing for very dense storage, and it is back-end-of-line compatible, allowing for on-chip fabrication to eliminate the von Neumann (off-chip memory access) bottleneck. Crossbar arrays connect many devices in parallel with individual devices read and written by selecting the appropriate word and bit lines. They are susceptible to large leakage currents due to the many parallel devices between non-selected and selected lines, and thus require a current limiting access device in series with every memory cell. Ovonic threshold switches (OTSs) are such devices. OTSs are highly resistive in the off state but switch to a highly conductive on state at some threshold voltage due to a field-dependent conduction mechanism in amorphous chalcogenides. They remain on while a minimum holding current passes through the device, then rapidly switch off. As access devices, OTSs must be designed such that all unselected devices remain off while a selected device passes enough current to read or write a memory cell regardless of the initial state of any memory cell in the array [1]. However, there is still debate on the mechanism(s) behind ovonic switching and thus on how a given design will behave. Recent models that explain ovonic switching include (i) field-based switching due to the filling of trap states near the fermi band in amorphous Ge₂Sb₂Te₅ (a-GST) [2], (ii) field-assisted crystallization, where crystals are stable in phase change materials and unstable in OTSs once the field is removed [3], and (iii) a field-assisted thermal phenomenon in doped a-GST [4]. We model ovonic switching in a-GST as a field-assisted thermal phenomenon by coupling field and temperature dependent conductivity terms [5], simulate switching in COMSOL Multiphysics, and find good agreement with experimental ovonic switching data for a-GST. We further simulate a phase change memory device in series with an OTS using our finite element phase change model [6]–[8].

[1] D. Kau et al., IEDM, (2009), pp. 1–4 DOI: 10.1109/IEDM.2009.5424263.
[2] D. Ielmini, Phys. Rev. B, 78, 3, p. 035308, Jul. (2008) DOI: 10.1103/PhysRevB.78.035308.
[3] M. Nardone et al., Appl. Phys. Lett., 94, 10, p. 103509, Mar. (2009) DOI: 10.1063/1.3100779.
[4] M. Le Gallo et al., J. Appl. Phys., 119, 2, (2016) DOI: 10.1063/1.4938532.
[5] A. Faraclas et al., IEEE Electron Device Lett., 32, 12, pp. 1737–1739, Dec. (2011) DOI: 10.1109/LED.2011.2168374.
[6] Z. Woods and A. Gokirmak, IEEE Trans. Electron Devices, 64, 11, pp. 4466–4471, Nov. (2017) DOI: 10.1109/TED.2017.2745506.
[7] J. Scoggin et al., Pre-print arxiv.org/abs/1810.07764 (2018).
[8] J. Scoggin et al., Appl. Phys. Lett., 112, 19, p. 193502, (2018) DOI: 10.1063/1.5025331.

Vanadium oxide shows versatile properties which can be triggered using temeprature as well as electric bias. Due to this, it has been proven to be a potential material in the field of resistive random access memory both as a memory element and selector. Here, we fabricate asymmetric cross-points devices of amorphous vanadium oxide films to analyse its resistive switching performance related to different oxide phases as well as phase-change induced insulator-to-metal transistion. Devices show a characteristic volatile threshold switching after electroforming at high voltages, whereas lower voltages reveal a forming-free non-volatile bipolar resistive switching behaviour. Volatile switching performance is apolar, stable, reversible, and symmetric for over 10³ consecutive cycles. While, non-volatile bipolar restive switching is asymmetric and stable over 200 cycles. Based on the achieved electrical characterisation and compositional analysis, it is suggested that the volatile threshold switching is due to the presence of VO₂ component, while non-volatile switching is due to the presence of V₂O₅ component in the mixed phase amorphous vanadium oxide film. This work highlights the ocuurance of bifunctional memory behvaior in assymetric cross-point devices and further provides guidelines to tailor the switching behaviour.

The electronic structure and conduction mechanism in chalcogenide-based Ovonic
threshold switches (OTS) used as selectors in cross-point memory arrays is derived from density
functional calculations and quasi-Fermi level models. The experimentally most favorable
compositions are found to lie in the 4:2 coordinated region of the phase stability maps of IV-VI
semiconductors. The switching mechanism in OTS is primarily electronic. It uses a specific
electronic structure, with a wide tail of localized states below the conduction band edge. In
amorphous GeSe2-x the valence band top consists of high effective mass Se pπ states with few band tails. The conduction band consists of Ge-Se antibonding states with low effective mass, and with a broad tail of localized Ge-Ge antibonding states below this band edge. This leads to the OTS behavior. At high fields the electron quasi-EF moves up through these tail states, lowering the conductivity activation energy, and giving the non-linear switching process. GeSe2-x is the most favorable OTS material because it has the optimum band gap.

The reversible material phase transition in chalcogenide compounds, known as phase-change memory (PCM) has been studied for half a century [1]. More recently, some PCM devices have been approaching the fundamental scaling limits of this technology [2]. This talk will present recent highlights from our research on PCM and related materials [3]. The results span from measurements of thermal and electrical properties of PCM devices and their interfaces, to understanding their fundamental size and energy limitations. For example, we show that the energy-efficiency of PCM programming can be significantly improved by reducing the pulse width (down to few nanoseconds). We use Raman thermometry and scanning thermal microscopy (SThM) to probe the temperature in functioning PCM devices [4,5], and uncover thermal and electrical contact resistance effects. We also introduce PCM devices which rely on emerging two-dimensional (2D) materials like graphene or MoTe₂, either as thermal insulation barriers [6] or as the switching material itself. Finally, we use carbon nanotubes (CNTs) as ultra-narrow electrodes to probe the insulator to metal (IMT) transition of VO₂ at the nanoscale. These results probe the fundamental limitations of PCM technology, as required to guide future energy-efficient designs.

References: [1] S. Ovshinsky, Phys. Rev. Lett. 21, 1450 (1968). [2] F. Xiong, A.D. Liao, D. Estrada, E. Pop, Science 332, 568 (2011). [3] F. Xiong, E. Pop et al., IEEE Intl. Electron Devices Meeting (IEDM), Dec 2016. [4] S. Fong, C. Neumann, H.-S.P. Wong, IEEE Trans. Electron Dev. 64, 4374 (2017). [5] E. Yalon, S. Deshmukh, M. Muñoz Rojo, F. Lian, C.M. Neumann, F. Xiong, E. Pop, Scientific Reports 7, 15360 (2017). [6] C. Ahn, C. Neumann, E. Pop, H.-S.P. Wong et al., Nano Letters 15, 6809-6814 (2015).

Phase change memory devices become practical for non-volatile storage at small dimensions due to reduced power and predictable device operation. In larger scale cells, devices can be locally melted due to filament formation and liquid filaments can be retained in parts of the cell for a long time even if most or all of the cells are initially amorphized during long fall-times. The complex amorphization and crystallization dynamics make these large cells more unpredictable and enable their applications as physically unclonable functions (PUF) [1,2].

Computational analysis of the complex amorphization-crystallization dynamics in phase change memory devices with large geometries is important to understand the evolution of phase distributions and temperature profiles during programming of these devices. In this work, we conduct electrothermal finite element simulations of reset operation on a large Ge₂Sb₂Te₅ (GST) cell using the framework we have developed in COMSOL multiphysics [3]-[9] and analyze the complex dynamics of amorphization, nucleation and growth during electrical stress. We input voltage waveforms measured from electrical characterization of on-oxide GST line cells with bottom metal contact pads and Si₃N₄ capping. A 2D polycrystalline model of the experimentally measured cells (~360 nm wide, ~400 nm long and ~50 nm thick) is constructed in the simulations. Access devices are modeled using the spice models. The simulations capture some of the interplay between changes in the device resistance due to heating and phase changes and current fluctuations.

References:
[1] N. Noor et al., Materials Research Society Fall Meeting & Exhibit, 2017.
[2] R. S. Khan et al., Materials Research Society Spring Meeting, 2017.
[3] Z. Woods and A. Gokirmak, IEEE Trans. Electron Devices 64, 4466 (2017).
[4] Z. Woods et al., IEEE Trans. Electron Devices 64, 4472 (2017).
[5] J. Scoggin et al., Applied Physics Letters 112, 193502 (2018).
[6] J. Scoggin et al., Pre-print arxiv.org/abs/1810.07764 (2018).
[7] A. Faraclas et al., IEEE Trans. Electron Devices 61, 372 (2014)
[8] A. Faraclas et al., IEEE Trans. Electron Devices 32, 1131 (2011)
[9] G. Bakan et al., Scientific reports 3, 2724 (2013)

Phase change memory (PCM) is one of the most promising non-volatile memory technologies that offers low-power, high-speed, and high-density operations. PCM utilizes the large resistivity contrast between amorphous and crystalline phases of chalcogenide materials like Ge₂Sb₂Te₅ (GST). The large resistivity contrast between the phases allows multilevel-cell (MLC) operation, which comes with some challenges including the upward resistance drift observed in PCM cells. A cell programmed to a certain state, if experiences resistance drift, may be inaccurately read some time after programming.
High-speed electrical characterization techniques allow us to monitor resistance drift behavior of the amorphous GST cells for a very short duration [1] and the slow DC measurements can be used to monitor the drift over a long period of time [2]. Our earlier resistance drift measurements were performed in 300 K to 675 K range [1]. In this work, we have amorphized GST line cells at cryogenic temperatures (down to 85 K) and monitored their resistance drift. We observe significant resistance drift of the amorphous GST at low temperatures. The viscosity of the material is expected to be very high to allow any significant structural relaxation at these low temperatures. Furthermore, we observe significant sensitivity to light: cells exposed to light for a period of time show a drop in resistance, and a recovery of the resistance after light is blocked. The time scale for these changes are in the order of seconds to minutes, suggesting that the changes in the resistance are predominantly due to emptying and filling of charge traps, and not due to thermal perturbations on the cells due to exposure to light.
Since we do not expect a rapid change in the number of charge traps via structural relaxation at cryogenic temperatures, our observations point to filling of the existing charge traps (charge trapping) as the main reason for upward resistance drift in phase change memory cells. Based on these results, we predict that it is possible to suppress resistance drift by increased capacitive coupling between the active region of PCM cells and the surrounding electrodes.

References:
[1] F. Dirisaglik et al., “High speed, high temperature electrical characterization of phase change materials: metastable phases, crystallization dynamics, and resistance drift,” Nanoscale, vol. 7, no. 40, pp. 16625–16630, Oct. 2015.
[2] N. Noor et al., “Short and Long Time Resistance Drift Measurement in Intermediate States of Ge₂Sb₂Te₅ Phase Change Memory Line Cells,” in Mat. Res. Soc. Spring Meeting, 2017.

Interfacial phase-change memory (iPCM) devices based on a Ge-Te/Sb-Te superlattice have lower switching energy than conventional phase-change memory due to optimization of the structure of the chalcogenide layers allowing minimization of the thermal losses associated with the switching process [1]. To date it has been widely believed that the structural changes responsible for switching in iPCM occur at the interfaces of the superlattice structure [2-4]. Recently, however, it has been proposed that resistive switching in iPCM may occur within a very narrow region of the superlattice device structure that has a specific atomic ordering. In the current work, the superlattice volume that is switched in iPCM was minimized by introducing a laterally connected contact to the superlattice film located between the two contacts perpendicular to the superlattice growth direction (van der Waals gaps plane in a superlattice structure). The resultant three-terminal iPCM device allowed the testing of selective regions of the device. This configuration is especially useful for the characterization of asymmetric iPCM devices and bipolar switching performance. The latter is believed to be related to the contact interfaces, which could be identified by the use of the third electrode. Previously, it was shown that iPCM devices based on Ge-Te/Sb-Te superlattices have a different switching thermal dependence than alloy-based phase-change memory [5]. In addition, multi-level and bipolar switching and external magnetic field effects [6] of iPCM devices resistance were found. The newly proposed device structure will allow for detailed characterization of these effects as well as the partial volume switching of iPCM.
In conclusion, the dependence of iPCM switching on the distribution of the applied switching electric field within the volume of Ge-Te/Sb-Te superlattices was studied with the use of a three-terminal device architecture. The effects of bipolar switching, high temperature operation and external magnetic field were assessed.
This work was supported by JST-CREST (JPMJCR14F1). A part of this study was supported by NIMS Nanofabrication Platform in Nanotechnology Platform Project sponsored by the Ministry of Education, Culture, Sports, Science and Technology (MEXT), Japan.
[1] R. E. Simpson et al., Nat. Nanotech., 6, 501-505 (2011).
[2] J. Tominaga et al., Adv. Mat. Interfaces, 1, 1300027 (2014).
[3] A. V. Kolobov et al., APL Materials, 2, 066101 (2014).
[4] X. Yu, J. Robertson, Sci. Rep., 5, 12612 (2015).
[5] K. V. Mitrofanov et. al., Jap. J. of Appl. Phys. 57, 04FE06 (2018).
[6] J. Tominaga et al., Adv. Funct. Mater., 27 (40), 1702243 (2017).

Phase change memory (PCM) is a non-volatile memory technology which utilizes resistance contrast of amorphous and crystalline phases of phase change materials such as Ge₂Sb₂Te₅ (GST). PCM cells are two-contact devices which offer fast read/write speed, high density, large endurance, and possibility of monolithic integration with CMOS at the back-end-of-the-line. There has been a growing interest in integration of PCM devices for computation in memory and neuromorphic computing.
It is possible to implement logic functions, multiplexing and data routing using phase change elements with multiple contacts [1], [2]. The logic functionality is achieved using two physical mechanisms that can be implemented at nanoscale: (i) isolation of some of the contacts from the others using local amorphization between other contact pairs, (ii) thermal cross-talk between the regions that are heated to be melted and the previously amorphized regions. The first allows for isolation of some of the input and the output terminals, and the second allows for toggle operations through crystallization of some regions while others are being amorphized. 6-contact phase-change devices can be implemented as toggle multiplexers or flip-flops when interfaced with 5 CMOS transistors, using ~50% CMOS footprint compared to the conventional CMOS alternatives with the added advantage of non-volatility. This approach can enable monolithic integration of 100s of GB of PCM atop CPUs for computer-on-chip applications.
We have performed a 2D computational analysis of six-contact phase change logic devices using our finite element framework that can simultaneously capture amorphization-crystallization dynamics including nucleation, growth, and grain boundary melting and electro-thermal phenomena in COMSOL Multiphysics [3]–[5]. Our analysis includes required write currents and pulse durations, maximum read voltages and read disturb, necessary access circuitry, design of device geometry and the thermal boundary conditions.

References:
[1] N. H. Kan’An, H. Silva, and A. Gokirmak, “Phase change router for nonvolatile logic,” in Device Research Conference - Conference Digest, DRC, 2014.
[2] N. Kanan, “Phase Change Devices for Nonvolatile Logic,” 2017.
[3] Z. Woods and A. Gokirmak, “Modeling of Phase-Change Memory: Nucleation, Growth, and Amorphization Dynamics During Set and Reset: Part I—Effective Media Approximation,” IEEE Trans. Electron Devices, vol. 64, no. 11, pp. 4466–4471, Nov. 2017.
[4] Z. Woods, J. Scoggin, A. Cywar, L. Adnane, and A. Gokirmak, “Modeling of Phase-Change Memory: Nucleation, Growth, and Amorphization Dynamics during Set and Reset: Part II - Discrete Grains,” IEEE Trans. Electron Devices, 2017.
[5] J. Scoggin, Z. Woods, H. Silva, and A. Gokirmak, “Modeling Heterogeneous Melting in Phase Change Memory Devices,” Oct. 2018.

Phase-change memory (PCM) is an industrial concerned technology that can be used both for traditional computer architecture and other new emerging architectures [1], such as in-memory computation. Generally, the transitions between order and disorder phases are employed to record data. However, one of the drawbacks of PCM is the relatively high power consumption. Therefore, controlling the transition without going through the amorphous phase in these materials is one of possible methods to reduce switching energy. In this talk, we will introduce two cases recently explored with DFT calculations for order-order transition for PCM materials. The first case is the stacking-fault motion induced Insulator-metal transition in GeTe/Sb₂Te₃ superlattice [2]. This transition may provide a significant change of carrier concentration and indicate a low energy-consumption process with a low energy barrier of atom motion for PCM. Secondly, our TDDFT molecular dynamics [3] reveals an unexpected effect of optical excitation in the experimentally observed rhombohedral-to-cubic transition of GeTe. The excitation induces coherent forces along [001], which is attributed to the unique energy landscape of Peierls distorted solids. The forces drive the A_1g optical phonon mode in which Ge and Te move out of phase. Upon damping of the mode, phase transition takes place in ps scale, which involves no atomic diffusion, defect formation, or the nucleation and growth of the cubic phase of GeTe.

[1] Xian-Bin Li et al, Phase-Change Superlattice Materials toward Low Power Consumption and High-Density Data Storage: Microscopic Picture, Working Principles, and Optimization, Adv. Funct. Mater. (2018) https://doi.org/10.1002/adfm.201803380.
[2] Nian-Ke Chen, Xian-Bin Li* et al., Metal-Insulator Transition of Ge-Sb-Te Superlattice: An Electron Counting Model Study, IEEE Trans. on Nanotech. 17, 140 (2018).
[3] Nian-Ke Chen, Xian-Bin Li* et al, Directional Forces by Momentumless Excitation and Order-to-Order Transition in Peierls-Distorted Solids: The Case of GeTe, Phys. Rev. Lett. 120, 185701 (2018).

Phase change materials (PCM) can exist in two different states, namely, as amorphous and crystalline solids, which can have markedly different physical properties. This is the basis for the usage of PCM for data storage applications. The ternary alloy of germanium, antimony and tellurium (GST) has been intensely studied, because at the stoichiometric composition Ge₂Sb₂Te₅, the amorphous phase can crystallize congruently (without change in composition) and rapidly. This makes it a material of choice for random access memory devices (pc-RAM). However, the amorphous phase spontaneously crystallizes at relatively low temperature (160 °C). For automotive applications, this temperature has to be increased, for example by enriching the stoichiometric GST with germanium [1]. However, then the crystallization is accompanied by a composition change: upon crystallization, the additional Ge segregates into the amorphous phase. This can lead to the nucleation and growth of a new crystalline phase which is rich in Ge. For the development of new materials for automotive applications devices, it is important to better understand this complex microstructure formation process.

Phase-field models are continuum models based on nonequilibrium thermodynamics [2,3]. Interfaces and surfaces are implicitly described by one or several scalar fields that often can be identified with order parameters. Their time evolution is naturally coupled to the transport of chemical components by diffusion. We have developed a grand-canonical multi-phase-field model for the crystallization of GST which takes into account three phases: the amorphous state, stoichiometric GST and a Ge-rich crystalline phase. We use a pseudo-binary approximation of the phase diagram and couple the phase fields to a single concentration field. The kinetic parameters of the model are determined from data on stoichiometric GST available in the literature [4]. As a first application, we study the isothermal crystallization of a homogeneous Ge-enriched amorphous GST film. The critical nuclei of the Ge-rich crystalline phase are computed, and the results are compared to classical nucleation theory. Growth of supercritical nuclei is then simulated, and the results are compared to recent experimental data [5].

[1] E. Palumbo, P. Zuliani, M. Borghi, R. Annunziata, Forming operations in Ge-rich Ge_xSb_yTe_z phase change memories, Solid-State Electronics 133 38-44 (2017)
[2] M. Plapp, Unified derivation of phase-field models for alloy solidification from a grand-potential functional, Phys. Rev. E 84, 031601 (2011).
[3] R. Folch and M. Plapp, Quantitative phase-field modeling of two-phase solidification, Phys. Rev. E 72, 011602 (2005).
[4] J. Orava, A.L. Greer, B. Gholipour, D. W. Hewak, C. E. Smith, Characterization of Supercooled Liquid Ge₂Sb₂Te₅ and its Crystallization by Ultrafast-heating Calorimetry, Nat. Mater. 2012, 11, 279 283
[5] M. Agati, F. Renaud, D. Benoit, A. Claverie, MRS Communications (2018), 1 of 8

Elemental antimony has been recently shown to be a suitable material for electronic phase change memories (PCM) when confined in ultrathin films [1].
Despite the extreme proneness to crystallization in the bulk, amorphous Sb has been demonstrated to be stable at 300 K for sufficiently long time to be used in PCM when it is grown in films as thin as 3 nm confined by capping layers [1]. The reasons behind the stabilization of the amorphous phase under these conditions are, however, unclear. Atomistic simulations can provide insights onto the several effects that might concur to stabilize the amorphous phase in confined geometry. To this end, we have generated an interatomic potential for antimony by using a machine learning approach, namely by fitting a large database of ab-initio energies by means of the neural network scheme introduced by Behler and Parrinello [2]. The same method was used previously to generate an interatomic potential for the phase change compound GeTe [3]. The potential has been validated by computing the structural properties of the crystalline, amorphous and liquid phases of bulk antimony. Then molecular dynamics simulations based on the neural network potential have been performed to investigate the crystallization kinetics of Sb in confined geometries. Possible effects on the crystallization kinetics induced by the presence of the capping layer in ultrathin films will be discussed.

[1] M. Salinga et al., Nat. Mater. 17, 681 (2018).
[2] J. Behler, and M. Parrinello, Phys. Rev. Lett. 98, 146401 (2007).
[3] G. C. Sosso, G. Miceli, S. Caravati, J. Behler, and M. Bernasconi, Phys. Rev. B 85, 174103 (2012).

The transport properties of Ge-Sb-Te (GST) ternary phase-change materials can be tuned by controlling their atomic structure and the concentration of charge carriers. Doping with atoms of different chemical species is a major method to reach this aim in GST crystals. The effects of doping on the transport properties of GST are challenging to study experimentally, since the crystalline phase of doped GST samples generally possesses a complicated microstructure, consisting of grains with different compositions. They are also challenging to investigate by first-principles methods based on the calculation of Kohn-Sham wave functions, as huge supercells are needed to describe the unavoidable chemical disorder among Ge, Sb, dopant atoms, and vacancies.

In this work, we perform first-principles calculations of the electronic structure and electrical conductivity of Si-doped GST fcc-crystals, using the spin polarized relativistic Korringa-Kohn-Rostoker (KKR) method based on the multiple-scattering theory [1,2]. The doped GST crystals have been described with a rock-salt unit cell, in which the chemical disorder is taken into account through the coherent potential approximation (CPA). The accuracy of the results is verified by comparing, for several Si contents, the density of electronic states (DOS) calculated with this method and with a method that uses Kohn-Sham wave functions in big supercells. The calculated Bloch spectral function (BSF) shows the dispersion of the electron states and its modification with doping. We show the Si-doping dependence of the electrical conductivity, and discuss it in terms of the concentration of charge carriers and of the modification of their scattering by the atomic disorder induced by doping. These results can be used to model samples, the microstructure of which consists of grains with different concentrations of Si atoms, each grain being described by a conductivity that depends on its composition.


References:
[1] H. Ebert et al. , The Munich SPR-KKR package , version 6.3, http://ebert.cup.uni-muenchen.de/sprkkr
[2] H. Ebert, D. Ködderitzsch, and J. Minár, Rep. Prog. Phys. 74 , 096501 (2011).

One of the more active research areas in phase-change memory is the concept of van der Waals (vdW)-bonded GeTe/Sb₂Te₃ superlattices (SL) also known as interfacial phase-change memory (iPCM). Devices based upon GeTe/Sb₂Te₃ SL structures have demonstrated both lower energy consumption as well as order of magnitude faster switching speeds. Unlike conventional Ge-Sb-Te alloy-based devices in which the SET and RESET states correspond to the crystalline and amorphous phases, the SET and RESET states in SL devices are both crystalline. Transmission electron microscopy studies of GeTe/Sb₂Te₃ SL structures structures have revealed that GeTe and Sb₂Te₃ blocks intermix in as-grown structures implying that the switching mechanism may be different than originally believed. In a recent work [1], it was proposed that the switching process is associated with the reconfiguration of vdW gaps along with concomitant deviations in the local stoichiometry from the GeTe/Sb₂Te₃ quasibinary alloys. This model offers an explanation of why the large conductivity difference between the SET and RESET states persists in the presence of intermixing. In the model, dynamic reconfigurations of the vdW gap in response to external stimuli such as heat or electric field are shown to lead to large changes in conductivity. In this work, we use density-functional theory to investigate the effects of electric field on a variety of proposal structures in the form of van der Waals blocks including the previously proposed Ferro, Kooi, Petrov, and Inverted-Petrov structures. Both 0K and finite temperature Molecular dynamics calculations were carried out using a supercell slab geometry for each structure with a vacuum gap taking advantage of the lack of dangling bonds on the outermost Te-plane that constitute one side of a vdW gap with an externally imposed sawtooth like potential that maintained periodic boundary conditions with the potential discontinuity confined to the vacuum region. The Ferro structure was found to undergo the largest structural rearrangement and the Kooi structure the least. Changes in local geometry and effective charges for each structure will be discussed. The effect of temperature will also be discussed via molecular dynamics studies using the same geometry.
This work was supported by JST-CREST (JPMJCR14F1).

[1] A. V. Kolobov, P. Fons, Y. Saito, and J. Tominaga. Atomic reconfiguration of van der Waals gaps as the key to switching in GeTe/Sb2Te3 superlattices. ACS Omega, 2(9):6223–6232, 2017.

As electrical device components are scaled down to atomic sizes, management of waste heat becomes a major issue in device performance. Minimizing energy consumption of electrical components and waste heat generation are critical issues in the operation of electronic devices devices. This has spurred interest in alternative cooling mechanisms, such as those provided by electrocaloric, magnetocaloric, elastocaloric, and thermoelectric materials, for mitigating issues associated with waste heat generation. In this work, we provide theoretical predictions for an alternative cooling mechanism, accomplished by utilizing electrostatic gating to induce structural phase transitions in monolayer materials. We refer to this mechanism as the electrostaticaloric effect in reference to the mechanism of electrostatic doping that drives the structural phase transformation and entropy change in the material. Recent predictions and experimental observation that electrostatic gating can induce structural phase transformations in monolayer materials opens the possibility for new application areas. Here, we explore the potential for electrostatically-induced structural phase transformations in monolayer MoTe₂ to be used in a Carnot refrigeration cycle. We predict that a temperature change of 10–15 K may be possible in devices that utilize monolayer MoTe₂ as the active phase change material. This mechanism may prove useful for future electrical devices which require cooling at the component level, and for which small monolayer devices are necessary.

Typical phase-change materials used in microelectronics rely either on amorphous-crystalline transitions or on redox reactions. Each of these mechanisms rely on slow and energy-intensive processes, such as heating and quenching or solid-state ion diffusion. Another approach without these drawbacks is to use materials with quantum phase transitions, but these are usually activated and maintained thermally rather than electrically. However, some reports exist of electrically-activated charge-density wave (CDW) transitions, which are similar to a Peierls distortion and sometimes observed in low-dimensional materials.

Inspired by work in selenide-based superlattices, where charge exchange between layers¹ can modulate an emergent CDW transition,² our group is working on synthesis and characterization of superlattices based on a well-known CDW material, TaS₂. Using the novel approach of depositing thin modulated layers of Ta and Sn that are combinatorially sputtered from metallic targets in the presence of sulfur plasma, we build libraries of such superlattices from the “bottom-up.” This is new because sulfide based superlattices have not been previously made using this approach. The deposition process is monitored using quartz crystal microbalances, which provide real-time in-situ feedback on the growth conditions.

Our x-ray reflectivity results show that we can form ultrasmooth thin films with control of the integer number of 2D monolayers in the repeating structure. We have also demonstrated that by altering deposition conditions we can vary sulfur content and adjust composition to promote formation of either SnS or SnS₂ layers, according to RBS measurements. Similar to selenium-based superlattices, changing layer thicknesses is expected to affect the electrical transport properties, since this changes the degree of charge transfer between the layers. Overall, the initial results from this work suggest that a similarly superlattice-modulated CDW phase transition may be observed in TaS₂-based superlattice materials.

1. Bauers, S. R., Merrill, D. R., Moore, D. B. & Johnson, D. C. Carrier dilution in TiSe 2 based intergrowth compounds for enhanced thermoelectric performance. J. Mater. Chem. C 3, 10451–10458 (2015).
2. Falmbigl, M., Fiedler, A., Atkins, R. E., Fischer, S. F. & Johnson, D. C. Suppressing a charge density wave by changing dimensionality in the ferecrystalline compounds ([SnSe]1.15)1(VSe2)n with n = 1, 2, 3, 4. Nano Lett. 15, 943–8 (2015).

The crystallization process of phase change materials has been largely investigated in literature in a wide range of temperature and for many compositions. On the contrary, there is a lack of information on the mechanisms governing the amorphization, since it can be obtained only by driven conditions, largely far from equilibrium, usually implying melting and quenching. Understanding which are the accessible local atomic configurations towards amorphization and how atoms move through a crystalline environment, even highly defective, hence, seems desirable. With this aim we have investigated the damaging process of crystalline GeSbTe alloys up to amorphization, by using ion irradiation. Irradiation with light ions (Ar⁺) has been adopted to produce diluted cascades, which allows very low amount of disorder to be introduced and to follow the structural evolution towards complete amorphization.
In situ reflectivity measurements and ex-situ resistance and Raman spectroscopy analysis have been employed to study the disordering process and its impact on the electrical properties and on the chemical bonds. Different compositions are compared: Ge₂Sb₂Te₅, GeSb₂Te₄ and GeTe. The effect of disordering is studied at low temperature (77K), at which the atomic mobility and disorder recovery is hindered, and at room temperature. For all the studied compositions large differences are observed upon irradiation at low or room temperature, indicating that there is a large effect of dynamic annealing. Several phase transitions are observed before amorphization. In all the cases, the first is the metal to insulator transition. However, the mechanism driving such a transition is different changing from the ternary compositions to the binary GeTe. Indeed, in the case of crystalline GeSbTe alloys (with Ge < 30%) the disordering process is dominated by the presence of the stoichiometric vacancies. The van der Waals like gaps present in the stable hexagonal crystalline structure act as preferential sink for the diffusion of the displaced atoms, and are therefore effective in stabilizing the disordered material. The filling of the gaps tunes the electronic and structural properties, driving the metal-insulator transition and the successive conversion into the metastable rocksalt structure and then in the amorphous.
In the case of GeTe there are no structural vacancy layers. The metal-insulator transition, observed at very low fluences, occurs in a mainly crystalline material with low resistivity value (~ 0.5 mOhm cm) and, according to Raman spectroscopy investigation, seems to be induced by Te vacancy, characterized by an energy level into the gap. As the irradiation fluence increases the displaced atoms form distorted octahedral bonds that, being energetically unfavourable, tend to disappear, going back to a more stable configuration. Such a process seems to be facilitated by the high atomic mobility that, according to crystal growth rate data available in literature, at room temperature is expected to be larger in GeTe than in GeSbTe alloys. As a consequence, GeTe exhibits unique marked self-healing properties and the irradiation fluence required for complete amorphization is one order of magnitude higher than that of GeSbTe alloys.
The role of the interfaces on the amorphization process has been also investigated and we found that by properly choosing the substrate and the capping layer it is possible to reduce or enhance the stability of the crystalline phases, in some cases inhibiting the amorphization process.
Our results indicate the possibility to tune the structural and electrical properties by interface and /or by defects engineering. Such an approach could help to avoid the employment of complex compositions, usually needed to finely tune the material properties, but which, on the other hand, can produce reliability problems, related to unwanted deviations from the optimized stoichiometry.

HfS₂ has attracted interests recently due to its distinctive properties such as high room-temperature mobility [1], finite bandgap [2], and high on-off current ratio [3]. The majority of the focus appears to be on in-plane properties. However, many exciting physical behaviors including the thermoelectric effect [4], field effect tunneling [5], and superconductivity [6] are tightly related to the c-axis (out-of-plane) electronic and thermal transport properties. However, the understanding of c-axis transport is still formative.
In recent X-ray and Raman spectroscopy measurements [7], an abnormal behavior was observed where a frequency increase – referred to as phonon “stiffening” - in the out-of-plane phonon mode A_1g accompanied an increase in temperature from 80K to 300K. This is in contrast to the normal behavior where increasing temperature leads to lattice expansion, thus resulting in decreasing phonon frequency or phonon “softening”. Beyond 300K, however, the trend in the Raman shift was found to reverse and phonon softening was observed. A correlated behavior also occurred in the thermal expansion coefficients (TECs) where a transition at 300K occurred between a region of low TECs at lower temperatures and high TECs at higher temperatures. To explain these observations, we performed first principles simulations using the Quasi-Harmonic Approximation (QHA) as well as first principles Car-Parrinello molecular dynamics (CPMD) simulations from 80K to 500K. The simulations revealed a fundamental change in the stacking pattern between adjacent Hf-S-Hf layers from AAA to ABC at 300K. The increasing temperature gave rise to a unique anharmonic mechanism. This was confirmed by comparing the calculated A_1g mode frequencies from temperature dependent phonon density of states with experimental data. Thus, we attribute the surprising phonon stiffening to a lattice phase transition that takes place at 300K where the transition is marked primarily by a change in stacking sequence. We also report for the first time a first principles fully-anharmonic temperature-dependent estimation of c-axis transport properties such as the carrier density, electron and phonon mean free path, mobility, and thermal conductivity.

References
[1]W. Zhang, Z. Huang, W. Zhang and Y. Li, "Two-dimensional semiconductors with possible high room temperature mobility," Nano Research, vol. 7, no. 12, pp. 1731-1737, 2014.
[2]C. Gong, H. Zhang, W. Wang, L. Colombo, R. M. Wallace and K. Cho, "Band alignment of two-dimensional transition metal dichalcogenides: Application in tunnel field effect transistors," Applied Physics Letters, vol. 103, no. 5, p. 053513, 2013.
[3]K. Xu, Z. Wang, F. Wang, Y. Huang, F. Wang, L. Yin, C. Jiang and H. He, "Ultrasensitive Phototransistors Based on Few-Layered HfS₂", Advanced Materials, vol. 27, no. 47, pp. 7881-7887, 2015.
[4]L.-D. Zhao, S.-H. Lo, Y. Zhang, H. Sun, G. Tan, C. Uher, C. Wolverton, V. P. Dravid and M. G. Kanatzidis, "Ultralow thermal conductivity and high thermoelectric figure of merit in SnSe crystals," Nature, vol. 508, no. 7496, p. 373, 2014.
[5]D. Sarkar, X. Xie, W. Liu, W. Cao, J. Kang, Y. Gong, S. Kraemer, P. M. Ajayan and K. Banerjee, "A subthermionic tunnel field-effect transistor with an atomically thin channel," Nature, vol. 526, no. 7571, p. 91, 2015.
[6]L. J. Li, E. C. T. O’Farrell, K. P. Loh, G. Eda, B. Özyilmaz and A. C. Neto, "Controlling many-body states by the electric-field effect in a two-dimensional material," Nature, vol. 529, no. 7585, pp. 185-189, 2016.
[7]S. Najmaei, M. R. Neupane, B. M. Nichols, R. A. Burke, A. L. Mazzoni, M. L. Chin, D. A. Rhodes, L. Balicas, A. D. Franklin and M. Dube, "Cross-Plane Carrier Transport in Van der Waals Layered Materials," Small, vol. 14, no. 20, p. 1703808, 2018.

As one of the most promising candidates for developing new neuromorphic architectures for non-von Neumann computing and information storage, Ge₂Sb₂Te₅(GST225) in photonic devices shows non-volatile, fast, and multi-step changes between the amorphous state and the meta-stable state. However, problems such as relatively low endurance and considerably high quenching power are still open challenges. In order to search for other composition ratios with desired properties, e.g. large difference in optical parameters, lower switching energy, etc., we have systematically investigated a broad composition range of the Ge-Sb-Te (GST) system. The GST combinatorial spread was fabricated by co-sputtering Ge, Sb and Te targets at room temperature. The composition ratios of sample segments characterized by wavelength dispersive spectroscopy confirmed that the composition variation across the Si wafer covers most part of the GST phase diagram. The structures characterized by synchrotron radiation, X-ray diffraction measurements and resistance mapping clearly show evolution of the phase-change temperature of the GST system: from a relatively low temperature (about 120 C°) in the Sb-Te rich region to a higher value (about 280 C°) in the Ge rich region. A phase boundary between the Sb-Te phase and the Ge-Sb-Te phase has been clearly identified using synchrotron diffraction. Refractive index and extinction coefficient were measured at each composition spot using a scanning ellipsometer: some composition ratios were found to have a small extinction coefficient in the amorphous state, a high phase-change temperature, and a large difference in the extinction coefficient between the amorphous state and the crystalline state, suggesting that such compositions can be promising for photonic devices for neuromorphic control. We find that compositions with promising properties are near the boundary between the Sb-Te and the Ge-Sb-Te phases, indicating that co-existence of the phases at microstructural level may be playing a role in giving rise to their properties. This work is funded by an ONR MURI (Award No. N00014-17-1-2661).

Besides graphene, many classes of 2D systems developed over the last few years: group-IV graphene-like Xenes (with X such as Si, Ge, Sn), transition metals dichalcogenides (such as MoS₂, WS₂, MoSe₂, WSe₂, MoTe₂), III-VI and V monochalcogenides (such as h-BN or InSe, GaS, GaSe, GaTe), and V-VI chalcogenides (V = As, Sb, Bi and VI = S, Se, Te).
Sb₂Te₃ and the pseudobinary alloy GeSbTe exhibit a 2D structure with the size of the building block depending on composition: Sb₂Te₃ (5 layers), GeSb₂Te₄ (7 layers), Ge₂Sb₂Te₅ (9 layers), Ge₃Sb₂Te₆ (11 layers), etc. The prerequisite for application, as well as for studying the fundamental properties, of such new crystalline materials is the capability of growing high-quality large-area films. Van der Waals (vdW) epitaxy, in which the epitaxy is mediated by weak vdW interactions, represents the way to realize such high quality films on large scale.
Despite the intense research efforts by many groups toward the vdW epitaxy, a general description of all the phenomena involved is not yet given. In fact, the epitaxial rules for 2D materials are really different from ordinary 3D materials. In 2D epitaxy, the prediction of the phase formed by the vdW heterostructures, the commensurability, the strain relaxation during the interface formation, and the role of the substrate surface reconstructions need to be investigated further.
In this contribution, a general description of the vdW epitaxy will be proposed.. We will discuss the epitaxial growth of the 2D GeSbTe alloy deposited by Molecular Beam Epitaxy on the low-mismatched InAs(111) substrates (lattice mismatch ≈1%) for both In- and As- terminated surfaces. The occurrence of the vdW epitaxy is discussed in comparison to other results from literature in order to identify the key mechanisms underlying the process.

Phase change materials (PCM) have attracted in the last years the interest of research as concrete candidates for the development of storage class memories. GeSbTe (GST) alloys along the GeTe-Sb₂Te₃ pseudo-binary line are nowadays used as the active material for non-volatile solid-state memories. [1]
For automotive applications, however, higher working temperatures are needed for embedded electronics. This prevents the use of Ge₂Sb₂Te₅ (GST225), which is among (GeTe)_m(Sb₂Te₃)_n compounds the one traditionally employed in PCM memory devices. One strategy to adapt PCM technology to the requirements of automotive applications is to tailor the crystallization temperature and thermal stability of GST by engineering the alloy composition. In this framework, Ge-rich GST alloys have shown increased crystallization temperatures proportional to the excess of Ge. [2]
Here we present a study of epitaxial GST films towards Ge-rich compositions grown by molecular beam epitaxy on Sb-passivated Si(111) substrates. Starting from stoichiometric GST225, the composition of Ge_2+xSb₂Te_5+x was controlled by increasing the atomic flux of Ge while keeping fixed the other growth parameters. The growth of Ge_2+xSb₂Te_5+x films was checked in-situ by reflection high-energy electron diffraction. We found a gradual loss of crystal quality with increasing Ge content, possibly related to the deficiency of Te which characterizes the present experiments. The evolution of the composition was studied by X-ray diffraction (XRD) radial scans and Raman spectroscopy. Interestingly, a combination of XRD rocking curves and azimuthal scans provides evidence of stronger in-plane rotational coherence with respect to the substrate, while the material is more prone to form domains misaligned out-of-plane. The characterization of the switching functionality will be carried out.

[1] S. Raoux, W. Welnic, and D. Ielmini, “Phase Change Materials and Their Application to Nonvolatile Memories,” Chem. Rev., vol. 110, no. 1, pp. 240–267, Jan. 2010.
[2] P. Zuliani, E. Varesi, E. Palumbo, M. Borghi, I. Tortorelli, D. Erbetta, G. D. Libera, N. Pessina, A. Gandolfo, C. Prelini, L. Ravazzi, and R. Annunziata, “Overcoming temperature limitations in phase change memories with optimized GexSbyTez,” IEEE Trans. Electron Devices, vol. 60, no. 12, pp. 4020–4026, 2013.

GST alloys have recently been reported to significantly increase the range of working temperatures of PCM memories, an important characteristic to extend their range of potential applications. Moreover, N doping may further increase their endurance and reduce the undesired drift of the low resistance state (crystalline state) upon time. However, the crystallization process i.e., the mechanisms by which amorphous N-doped Ge-rich GST transforms into one or several stable crystalline compounds, is rather unknown.
For this reason in this work we report on a systematic study of the structural and compositional changes which affect thin films of amorphous N-doped Ge-rich GST during thermal annealing at various temperatures. The structural evolution was studied using a variety of transmission electron microscopy techniques (HRTEM, SAED, ASTAR), both in-situ and ex-situ, and X-ray diffraction. The chemical composition was analyzed by STEM-HAADF and STEM-EDX.
We show that, during annealing at 380°C, the initially homogeneous Ge-rich GST material locally fluctuates in composition until phase separation occurs while no crystal nuclei forms yet. Then, only at 400°C, Ge crystals first form in the Ge-rich regions, followed by the nucleation of pure GST225 grains in the Ge-depleted regions. Annealing at higher temperatures allows the growth and coalescence of these grains until the film is totally (poly)crystalline. Finally, our study demonstrates that the crystallization kinetics of Ge-rich GST is limited by atomic (Ge) diffusion. Moreover, in the « crystalline » (set) state, the material consists of a mixture of pure Ge and GST225 grains of nanometric dimensions, a characteristic to keep in mind when studying the electrical behavior of PCM memories based on such materials.