Symposium SB04—Materials and Algorithms for Neuromorphic Computing and Adaptive Bio-Interfacing, Sensing and Actuation

Neuromorphic devices represent a powerful platform to create artificial neural networks, which will offer the chance to improve implantable devices towards the replacement of brain injuries. As widely reported in literature, the dominant synaptic communication is based on electrochemical mechanisms by releasing neurotransmitters [1]. However most of the existing implantable prosthetics are designed to operate mainly through intrinsic potentials of the brain, overlooking the chemical aspect. In this scenario, neuromorphic devices offer a unique new perspective: these devices can “learn” by receiving stimuli, changing their plasticity similar to physiological synaptic plasticity [2]. Recently we demonstrated a biohybrid synapse that formed a bidirectional coupling between PC12 neuron-like cells and an organic neuromorphic device. Applying a positive bias at the gate leads to oxidation of dopamine secreted from cells and the release positive ions in the electrolyte solution as well as electrons donated to the PEDOT:PSS channel. These processes de-dope the material resulting in a permanent change of resistance of the channel. Additionally, this device shows a high state retention after the pulses are stopped resembling long-term plasticity [3]. This work opens the pathway towards combining artificial neuromorphic systems with biological neural networks for future brain implants. Here we show a further expansion of these biohybrid synapses towards highly reliable prosthetics. Oxidized dopamine gradually forms a layer of polydopamine on the interface of cells and the device, which may result in blockage and limit the device sensitivity. Moreover, in case of prosthetic implants the device must be highly biocompatible and flexible. To tackle these challenges, we introduce a new technique of fabricating flexible organic electrochemical devices on highly biocompatible flexible substrates. These devices use polydopamine as a solid state electrolyte and simultaneously act as a de-doping agent for the PEDOT:PSS layer to increase the resistance of the channel for neuromorphic memory. A more resistive channel avoids excessive current through the device which decreases the energy consumption. The presence of amine in the PEDOT:PSS channel also reduces the diffusion of amines into the electrolyte leading to more stability and higher state retention. Furthermore, using polydopamine as the solid electrolyte significantly increases the energy barrier for ion/proton diffusion which eliminates the need for an access device or resistor. This will significantly increase the ease of design, fabrication and integration of these neuromorphic devices in biological environments. Polydopamine as a biomaterial melanin is inherently biocompatible, which is essential for viable bioelectronic material [4]. We expect these new flexible biocompatible ENODes can be used to bridge hybrid synaptic connections and could further be utilized in an array fashion for integration of future brain-machine interfaces.

References
[1] M. E. Burns and G. J. Augustine, “Synaptic structure and function: Dynamic organization yields architectural precision,” Cell, vol. 83, no. 2, pp. 187–194, Oct. 1995.
[2] J. J. Jun et al., “Fully integrated silicon probes for high-density recording of neural activity,” Nature, vol. 551, no. 7679, pp. 232–236, 08 2017.
[3] S. T. Keene, C. Lubrano, S. Kazemzadeh, A. Melianas, Y. Tuchman, G. Polino, P. Scognamiglio, L. Cinà, A. Salleo, Y. van de Burgt and Francesca Santoro. “A biohybrid synapse with neurotransmitter-mediated plasticity”, Nature Materials, VOL 19, 2020, 969–973.
[4]. Lee, H., M.Dellatore, S., M.Miller, W. & B.Messersmith, P. Mussel-Inspired Surface Chemistry for Multifunctional Coatings. J. Chem. Inf. Model. 318, 1689–1699 (2007).

The brain can perform massively parallel information processing while consuming only ~1- 100 fJ per synaptic event. I will describe an electrochemical neuromorphic device that switches at low energy (~80 fJ), and displays a large number of distinct, non-volatile conductance states within a ~1 V operating range. The tunable resistance behaves very linearly, allowing blind updates in a neural network when operated with the proper access device. These devices also display outstanding endurance achieving over 10⁹ switching events with very little degradation. I will describe our recent efforts at scaling and materials selection, allowing us to reach 20 ns write pulses and operation at high temperature (up to 120°C). In particular, we developed a fully lithographic process that allowed us to demonstrate sub-µm channel devices, opening the door to integration with Si driving circuitry. By carefully deuterating the electrolytes, we provide strong evidence that the secret to the high speed and low energy switching properties of these artificial synapses is the combination of electronic and protonic transport. Finally, we demonstrate that the working mechanism is quite general by fabricating and operating high-performance synapses based on MXenes. This generality is promising in terms of monolithic integration as MXenes can be chosen to be BEOL compatible with Si.

It becomes increasingly difficult to improve the speed-energy efficiency of traditional digital processors because of limitations in transistor scaling and the von Neumann architecture. Computing systems augmented with emerging devices such as memristors offer an attractive solution. A memristor, also known as a resistance switch, is an electronic device whose internal resistance state is dependent on the history of the current or voltage it has experienced. Because of its small size and fast switching speed, a memristor consumes a small amount of energy to update the internal state (training). Within large-scale crossbar arrays, memristors perform in-memory computing by utilizing physical laws, such as Ohm's law for multiplication and Kirchhoff's current law for accumulation. The current readout (inference) at all columns is finished simultaneously regardless of the array size, offering massive parallelism in vector-matrix multiplication and accelerating neural computing. This talk will first introduce an analog memristor that meets most requirements for in-memory computing in artificial neural networks. We will then showcase the integration of large-scale crossbar arrays of these memristors and the implementation of multilayer neural networks for several machine learning applications.

In-memory computing with combined logic and memory functionality is a promising energy-efficient solution for data-intensive processes like machine learning. A crucial component towards in-memory computing is the analogue nonvolatile resistive memory element. Two-terminal filamentary valence-change memory (VCM) based on metal oxides are extremely difficult to control due to the formation of a nanosized filament. On the other hand, more predictable and deterministic solutions like electrochemical random-access memory (ECRAM) based on the intercalation of lithium and proton ions have poor state retention when scaled to nanosized dimensions, and often utilize materials that are difficult to integrate into CMOS processes. A memory solution possessing all of these requirements have been elusive.

In this work, we combine the VCM and ECRAM concepts by replacing the highly mobile lithium and proton ions in ECRAM with less mobile oxygen vacancies. As a three-terminal ECRAM cell with a solid electrolyte, it possesses highly desireable linear and deterministic analogue switching at moderately elevated temperatures (~100-150C). On the other hand, the low mobility of oxygen vacancies at room temperatures provide nonvolatile state retention for days and weeks at scale without a selector, while also enabling the use of CMOS-compatible transition metal oxides. We further show how the switching and retention time are related to fundamental material properties like ionic resistance and chemical capacitance, and how changes in material processing conditions can be used to control the switching and retention properties.

In recent years, many novel computing paradigms have emerged that promise to complement the classic von-Neumann architecture, such as artificial neural networks and computation-in-memory. It is expected that these will enhance the functionality of everyday devices, such as in edge-AI and the internet of things as well as in datacenters for green-IT.
Analog accelerators based on crossbar arrays of non-volatile memristive devices (such as resistive switching ReRAM) are believed to be well suited for many of these applications, as they offer efficient (i.e., fast parallel and low power) operation and high integration densities. However, because of their high write energy and limited write count, applications of interest are limited to read-dominated ones, such as ANN inference or database queries. Yet, the effect of ReRAM device reliability issues as variability and drift on the crossbar operation needs to be explored.
To characterize the reliability of the crossbar, a representative computation is needed. The vector dot-product was chosen here because it can serve as a computational primitive for constructing more complicated operations such as matrix or tensor multiplications, which are among the most utilized operations in neural networks. Therefore, it can reveal the key behaviors, while still being sufficiently simple to allow for easy interpretation of the results.
We fabricated crossbar structures based on a ZrO₂/Ta ReRAM device stack. These feature a number of cells connected with one electrode to a common bitline, and can be viewed as a column of a passive memory array. The measurements presented were performed on 8-bit words stored in these. To perform a computation, an input voltage vector is applied to the block, the sum of the currents through all memristive devices is then the calculated dot-product.
Using this configuration, we experimentally demonstrated the feasibility of the operation. From these experiments, a number of important conclusions could be drawn regarding the memristive device characteristics. For example, it was found (using binary weights) that the spread in current levels depends strongly both on the accuracy of the programmed low resistance state value and on the resistance level of the high resistance state.
Furthermore, we identified typical failure mechanisms, such as read disturb, which is a change or drift in the device resistance caused by repeated reading of the cells or random bit-flips. We show how these are influenced by the magnitude of the applied voltages, and compare the behavior to that expected from the characterization of single cells. We demonstrate how they can be understood based on the material properties, such as the non-linear device kinetics being a key driver of low resistance state degradation. This allows us to identify the best operating conditions.
In a real world application the crossbar does not operate as a standalone device, but in conjunction with CMOS periphery. This periphery is needed to convert between the analog domain of the crossbar and the digital domain of the rest of the system. The impact of the measured device level failures on this interfacing between the analog and digital world is investigated through simulation. For these, we fitted our physics based compact model JART VCM v1b to measured data from the cells, which is able to describe the device-to-device and cycle-to-cycle variability, as well as mechanisms such as read disturb and random telegraph noise.

As our understanding of the brain’s physiology and pathology progresses, increasingly sophisticated technologies are required to advance discoveries in neuroscience and develop more effective approaches to treating neuropsychiatric disease. To facilitate clinical translation of advanced materials, devices, and technologies, all components of bioelectronic devices have to be considered. Organic electronics offer a unique approach to device design, due to their mixed ionic/electronic conduction, mechanical flexibility, enhanced biocompatibility, and capability for drug delivery. We design, develop, and characterize conformable organic electronic devices based on conducting polymer-based electrodes, particulate electronic composites, high-performance transistors, conformable integrated circuits, and ion-based data communication. These devices facilitate large-scale neurophysiology experiments and have led to discovery of a novel cortical oscillation involved in memory consolidation as well as elucidated patterns of neural network maturation in the developing brain. The biocompatibility of the devices also allowed intra-operative recording from patients undergoing epilepsy and deep brain stimulation surgeries, highlighting the translational capacity of this class of neural interface devices. In parallel, we are developing the high-speed conformable implantable integrated circuits and embedded acquisition and storage systems required to make high channel count, chronic neurophysiological recording from animals and human subjects possible. This multidisciplinary approach will enable the development of new devices based on organic electronics, with broad applicability to the understanding of physiologic and pathologic network activity, control of brain-machine interfaces, and therapeutic closed-loop devices.

By understanding the self-healing properties of biological systems it may be possible to build robust artificial systems with similar stable operation. Here we discuss how the nervous system of the Hydra - a millimeter-sized aquatic cnidarian - can reorganize itself in a matter of hours to regain function after half the neurons are lost. By studying how this simple nervous system confined to two layers of cells achieves this robust control of the body we hope to reveal ways we may be able to design and build robust, soft, biological neural control systems.

In animals, and especially humans, trained neural networks form associative learning patterns that allow rapid classification and decision making (e.g. object identification via touch sensing). In contrast, current engineered systems are unable to sense and classify information without significant instrumentation and software. But this might be possible in hardware alone if a system were constructed in a way that combines materials for distributed sensing, intrinsic computing (for classification), and memory (for learning). Only a handful of research works by Bao et.al, Lee et.al, Zhu et.al have tried to explore the artificial sensory nervous system, mostly using transistors, with just one work by Zhang et.al used threshold dependent switching devices for data processing.
Recently, Arrieta, et al demonstrated thin, flexible dimpled sheets, comprised of planar arrays of bi-stable domes, which reconfigure their global shapes in emergent forms when specific domes are inverted. The global shape of the sheet is dictated by both the local pattern and sequence of dome inversions. We propose to leverage the complex mechanics of these dimpled sheets combined with polymeric thin film non-volatile memristors to provide a novel foundation for combining autonomous reconfiguration with distributing sensing and embedded memory—akin to a peripheral nervous system—for onboard classification in engineered structures.
In this study, we demonstrate that dome inversion events can be sensed and remembered by integrating resistive strain sensors connected electrically to memory storage elements. The former is achieved by printing conductive traces into the dimpled sheets at positions near the individual domes, while the latter consists of polymeric thin-films that function as memory resistors. When the two are connected in series and biased with voltage, a transient change in sensor resistance caused by deformation of the dome encodes a permanent change in the resistance of the memristor. Through fabrication and experiments, we study the design and integration of both the sensor and memory storage elements for effective operation. An amplification circuit is used to magnify the changes in sensor resistance occurring during dome deformation, which allows the memristor voltage to eclipse the switching threshold. Using 1 sensor:1 memristor pairs, we demonstrate spatially distributed transduction and memory of dome inversions, and by applying a pulsatile voltage, we demonstrate incremental changes in the state of the memristor, which allows for temporal encoding of dome inversion events. These capabilities set the stage for an era of smart sensing and memory (via memristors) applications with emerging systems (reconfigurable sheet) that interact with the world for autonomous classification of tactile information.

Early detection of malign patterns in patients’ biological signals can save millions of lives. In this regard, artificial intelligence and machine learning can be used to detect patterns with super-human abilities. However, the practical clinical application of these methods is mostly constrained to an offline evaluation of the patients’ data. In-vivo and real-time use of machine learning is so far only a vision, hampered by the difficult integration of traditional electronics and dedicated hardware into the human body.
Previous studies have identified organic electrochemical transistors (OECTs) as ideal candidates in bioelectronics for biosignal monitoring and sensing. However, using them for pattern recognition in real-time has so far not been achieved. Likewise, OECT-based artificial synapses have been proposed, but never integrated into networks.
In the present study, we produce and characterize brain-inspired random networks composed of OECTs and employ them for time-series predictions and classification tasks using the machine learning framework known as reservoir computing.
Due to their biocompatibility and low energy consumption, these networks could serve in implantable devices to carry out real-time on-chip computation. We successfully pattern these networks on flexible and biocomptaible substrates and we apply them to learning tasks with classical databases (classification of flowers and time-series predictions).
Furthermore, we show the potential use of the networks for biofluid monitoring and biosignal analysis. For this purpose, we classify 4 classes of arrhythmic heartbeats with an accuracy of 88%. Each measurement is performed in PBS solution, in order to mimic the interface between the polymer and the biological cellular environment.
The results of this study introduce a novel paradigm for biocompatible computational platforms and may enable the development of ultra-low power consumption hardware-based artificial neural networks capable of interacting with body fluids, sensing chemicals, and monitoring biological tissues.

In biological neural networks communication relies on electric signals in the form of action potentials (APs) that trigger the release of neurotransmitters at the synapses, thus influencing the mechanisms of synaptic plasticity that regulates over time the transmission efficiency across the synaptic cleft. Neuromorphic computing and its progress intrinsically depend on faithfully replicating the communication mechanisms of biological neural networks. As such, electrical to chemical signal transduction is essential to fully mimic brain functions and establish in future an active interaction with biological tissues¹. The potential transition from fully artificial to bio-hybrid, integrated synapses is only enabled by neuromorphic systems that display synaptic conditioning based on biochemical signalling activity which typically involves more than one neuromodulator.
The conductive polymer mixture PEDOT: PSS can be exploited for the electrochemical detection of electroactive neurotransmitters² (i.e. dopamine, serotonin, ascorbic acid) through redox reactions while providing a suitable biocompatible interface with cells and neurons³.
Here, we report the first demonstration of a neuromorphic platform recapitulating the synergetic concourse of multiple neurotransmitters (NTs) that control synaptic weight modulation. This novel symmetric electrochemical neuromorphic organic device (SENODe) is capable of identifying three neurotransmitters that are simultaneously present and through modular microfluidic channels to mimic the recycling machinery of the synaptic cleft and control the output to replicate exocytosis-driven synaptic potentiation and depression⁴.Hence, the platform allows to transduce the neurotransmitters-mediated chemical signal, eliciting an output that is stored and read as a change in the polymer’s conductance, a reversible neuromorphic phenomenon that replicates neuronal’s short-term and long-term plasticity⁵.Moreover, by exploiting PEDOT:PSS electrochromism, the system’s readout is alternatively monitored as a variation of the polymer’s transmittance. The combination of the two readout mechanisms allows to emulate high order biological processes such as intrinsic forgetting, memory consolidation and NTs co-modulation, while also classic conditioning processes are replicated. Hence, the possibility to combine different NTs and input stimuli enableto program the synapse weights opens the way to the integration of array of devices which recreate complex, more refined brain’s learning processes.

1. Markram, H. A history of spike-timing-dependent plasticity. Front. Synaptic Neurosci. 3, (2011).
2. Gualandi, I. et al. Textile Organic Electrochemical Transistors as a Platform for Wearable Biosensors. Sci. Rep. 6, 33637 (2016).
3. Asplund, M. et al. Toxicity evaluation of PEDOT/biomolecular composites intended for neural communication electrodes. Biomed. Mater. 4, 045009 (2009).
4. Gu, C., Larsson, A. & Ewing, A. G. Plasticity in exocytosis revealed through the effects of repetitive stimuli affect the content of nanometer vesicles and the fraction of transmitter released. Proc. Natl. Acad. Sci. 116, 21409–21415 (2019).
5. Abbott, L. F. & Regehr, W. G. Synaptic computation. Nature 431, 796–803 (2004).

Neuromorphic electronics is a promising approach to enable dedicated computing at the site of sensing and actuation at the same time it offers new strategies to perform signal processing. Neuromorphic devices and systems combine signal processing and memory functionality in a manner to enable “training” and “evolution” of hardware to improve its performance for identification, perception and/or quantifying data and information. Our approach takes use of operating organic electrochemical devices and systems in electrolyte media containing electroactive monomers. This approach allows us to evolve the actual physical dimensions of devices while operating and performing signal processing. Organic electronic materials are unique in the sense that they can operate and be manufactured at the same conductions, which here allow us to make novel neuromorphic technology that relies on evolving critical device parameters of organic electrochemical transistors. We will report on device architectures, manufacturing approaches, materials and system design for organic evolvable electronics.

Ions are the primary charge carrier in physiological environments. Conventional inorganic electronic materials used in bioelectronic devices rely on surface capacitance for conversion of ionic potentials to electrons at the tissue interface for electrophysiological sensing and stimulation. In contrast, conducting polymers have mixed ion and electron conductivity that enables ion-based modulation of electronic charge carriers. However, it is not possible to individually control the properties of these polymers’ ion and electron conducting domains. Mixed-conducting particulate composites are an emerging material that allows precise tuning of ionic and electronic conduction states by varying particle size and density in a scaffolding matrix. We hypothesized that a similar strategy could be used to create an anisotropic ion conductor by controlling the particle path (ionic mean free path) length to allow ion conductivity only within the lateral plane. To test our hypothesis, we synthesized anisotropic ion conducting films composed of biocompatible ion conducting particles and organic dielectric materials. Electrochemical impedance spectroscopy and scanning electron microscopy measurements were performed to characterize the degree of anisotropy and mean free path of particles. We showed that such composites have uniformly distributed particles, hydration stability over weeks, scalable thickness, tunable out-of-plane conductivity, and consistent anisotropy for various particle and scaffold materials. Additionally, an anisotropic ion conductor can be designed as a soft, stretchable, and conformable transducer that directly interfaces with the skin for neural interface applications. We demonstrated the application of this material by acquiring biological signals via electrocardiography. Furthermore, we applied this ion membrane in organic electrochemical transistors (OECTs) to eliminate crosstalk between transistors, which is conventionally caused by the use of a shared electrolyte. This enables the fabrication of high-density transistor arrays, opening the avenue for organic integrated circuits and OECT-based neuromorphic computing devices. The development of this novel anisotropic ion conducting film has a wide range of implications and potential applications, ranging from anisotropic interfaces with biological tissue for high-resolution electrophysiology recordings and large-scale integration of ionic circuits, to directional control and programming of synaptic weights for artificial neural network and neuromorphic devices.

Biomembrane-based devices exhibit unique capabilities in the realm of neuromorphic applications. Doping model lipid membranes with voltage-active species transforms them into low voltage volatile memristors that dynamically and nonlinearly process incoming signals in a manner consistent with sensory synapses in the retina and olfactory systems. Unlike the brain however, volatile memristor networks do not have a built-in method to interpret processed signals. Reservoir computing (RC) may be one such way to interpret the nonlinear, history-dependent outputs from such networks. In the simplest terms, a reservoir is a network of nodes with inputs and outputs whose connection topology and weights do not change. Classification can be performed on reservoir outputs using supervised learning techniques to train a “readout layer” of artificial neurons, independent of the internal connectivity that defines the reservoir itself. Recently, volatile, solid-state memristors have been successfully used to create a reservoir for enbling image classification and chaotic temporal forecasting tasks. Low voltage, organic, and biocompatible materials, including those that more faithfully emulate the short-term plasticity of biological synapses, however, have not yet reported these capabilities but have the necessary properties to show similar success.

Herein, we propose the use of monazomycin-doped droplet interface bilayers (MzDIB) as candidates for reservoir computing nodes. We hypothesize that this material composition has unique properties that can reduce the overall size of the reservoir which will reduce footprint and energy expenditure. Using physical modeling, numerical simulation, and image classification tasks, we show that MzDIBs can accurately classify handwritten numbers digitized as small (5x5) or large (20x22) pixel representations of digits with appropriately sized reservoirs. A custom 5x5 pixel digit dataset consisting of ten distinct representations of each integer 0-9 is tested as proof of concept, with the larger 20x22 images consisting of the MNIST handwritten digit dataset with black borders removed. To process images using voltage-activated MzDIBs, pixel rows of image are converted into time varying voltage inputs fed at different rates into the reservoir. The voltages cause bi-directional, time-varying changes in membrane conductance due to Mz channel formation, such that the simulated final conductance states are then used as inputs to a 196x10 readout layer for image classification. We show that classification using a MzDIB reservoir network is feasible. Fully realized arrays with more complex connectivity have the potential to be hardware based reservoir computers due to their low operation voltage, biocompatibility, and ease of assembly.

Filamentary adaptive oxides operate on nanosecond time scales, with local temperatures as high as 1000 K, yet the impact of the thermal environment on their analog operation is often ignored. These devices are a promising new technology for neuromorphic and biomimetic computing applications, which require linear, analog resistance changes. Filamentary oxides can be set to a continuum of resistance states based on the amount of oxygen ions within a nanosized area, called a filament. If the maximum current is well controlled, the oxygen deficient region can be partially re-oxidized by reversing the bias, thus increasing the resistance of the device. The resistance is modulated by applying voltage pulses for times varying many orders of magnitude, from seconds to nanoseconds. The ability to realize these devices in neuromorphic computing models such as spike timing dependent plasticity or pattern recognition relies on achieving a repeatable, linear resistance change. Understanding the role that the thermal environment has on the distribution of oxygen ions will inform material choices for the electrodes, substrates, and packaging of neural network chips.

The main mechanisms that drive the oxygen ions in and out of the filament are drift, diffusion, and thermophoresis [1]; all of which are heavily influenced by the temperature. Temperature gradients are the driving force for thermophoresis, which is very strong in these devices due to a localized high temperature region where all the current passes. Previous work showed changing the thermal conductivity of the substrate by an order of magnitude, from 138 W/m-K to 1.2 W/m-K, increased the percent change in the resistance of a single pulse from 150% to 450% [2]. It was determined that when the substrate thermal conductivity was low, the electrode acted as a heat sink. To influence the thermal environment of the switching oxide more directly, this work focuses on changing the thermal properties of the electrodes while maintaining the same chemical environment near the filament. Previous studies have explored the impact of the electrodes in direct contact with the filament [3], however, this layer also alters the bonding and release of oxygen ions, and thus these experiments did not independently change the thermal environment. In this study, the thermal environment imposed by the electrodes will be independently changed by surrounding the filamentary HfO_x active layer with thin, 5 nm gold layers. The bulk of the electrodes will then be varied between silver, platinum, and TiN having thermal conductivities of ~429, 72, and 25 W/m-K. Initial finite element modeling suggests that increasing the thermal conductivity of the electrodes reduces the initial resistance change to the first pulse. This is important because these devices typically have a large initial resistance change that quickly levels off, instead of the desired linear change with the number of pulses. The digital, analog, and operation temperature of these devices are compared and supported by modelling work. By tailoring the thermal environment of the device, the temperature gradient is manipulated thus altering relative magnitudes of three main thermal driving forces, drift, diffusion, and thermophoresis. Fundamentally understanding how the electrode materials influence the local temperature distribution will lead to more reliable and repeatable filamentary devices for neuromorphic applications.

This work was supported by the AFOSR MURI under Award No. FA9550-18-1-0024. In part at the Georgia Tech Institute for Electronics and Nanotechnology, a member of the NNCI, which is supported by the NSF (ECCS-2025462). This material is based upon the work supported by the NSF GRFP under Grant No. DGE-1650044.

[1] D. G. Pahinkar, et al. AIP Advances 10, 035127 (2020)
[2] M. P. West, et al. Applied Physics Letters 116, 063504 (2020)
[3] W. Wu, et al. IEEE Electron Device Letters 38, 1019-1022 (2017)

Towards future computing and nonvolatile memory (NVM), memristors have gained wide attention for their in-memory computing and neuromorphic computing applications [1, 2]. The resistive switching (RS) dynamics in memristors rely critically on oxygen ions and their recombination with oxygen vacancies to partially rupture and form the conductive filament (CF) [3] for modulating the device between the low resistance state (LRS) and high resistance state (HRS).

In this study, we investigated the effect of oxygen ion concentration at the HfO_x/Ti interface on set and reset operation, LRS, and HRS characteristics in HfO_x-based filamentary memristors. Devices with higher oxygen ion concentrations at the interface were fabricated by performing oxygen plasma treatment at varying times on the as-grown ~5 nm HfOx active layer via atomic layer deposition (ALD). Detailed x-ray photoelectron spectroscopy (XPS) analysis and its depth profile results verify the excess oxygen ion present at the HfO_x/Ti interface, and thin TiO_x layer has been formed between the HfO_x and Ti layers, resulting in a HfO_x/TiO_x/Ti structure. This change in interface chemistry at the active layer and capping layer near the top electrode have impacted the set transition behavior more gradual, favorable for analog logic computation. This work demonstrates the importance of oxygen ion concentration and motion in controlled growth of CF during memristor set operation.

[1] J. J. Yang et al., Nature Nanotechnology, 3(7), 429–433 (2008).
[2] J. J. Yang et al., Nature Nanotechnology, 8(1), 13–24 (2013).
[3] Y. Li et al., J. of Physics D: Applied Physics, 51, 503002 (2018).

For more than a decade, a shift of paradigm has taken place within neural implant engineering with the idea of implementing active components and bringing out modalities capable of modulating the biological surroundings. Examples span from integration of micro-LED for photostimulation of cells, to microfluidic systems for local drug delivery.

This shift has been a cornerstone in neuroengineering that allowed to move from a passive monitoring to a dynamic modulation of the nervous system. Nonetheless, one of the missing parts is the ability to emulate certain functions of the neural system, such as neuronal connectivity, network plasticity or local computing. Indeed, to achieve a biologically relevant modulation, it is crucial to mimic as closely as possible neuronal processes happening in the body. The next generation of devices needs thus to focus on implementing biomemetic aspects to not only interface with the nervous system, but recreate these biological functions to locally sense, process and modulate it.
The central nervous system is at the essence of our ability to learn, memorize and compute information and at the core of this intricate and complicated network lays the synapse and its capacity of passing information from one neuron to another. The first building blocks to recreate the functions of this key element have already been developed with a non-volatile electrochemical neuromorphic organic device (ENODe)¹ which showed short- and long-term plasticity, laying down the foundation of neuromorphic devices for bio applications. Another important milestone has been the modulation of a biohybrid synapse through the inhibitory neurotransmitter dopamine released by neuron-like cells².
In this work, we propose to further develop these technologies and to create an artificial synapse capable of creating neuronal micro-circuits. In its nature, the synapse can strengthen or weaken over time and different mechanisms cooperate to achieve this plasticity, such as the release of neurotransmitters by the presynaptic neuron or the time difference of firing between the pre- and postsynaptic neurons. We are therefore investigating the modulation of an artificial synapse trough the presynaptic release of inhibitory or excitatory neurotransmitters and through a spike-time dependent plasticity (STDP) phenomenon. An unconnected pair of neurons is coupled together thanks to this artificial synapse showing a plasticity at the chemical level, with the neurotransmitters, and in the time domain, with the STDP, allowing a more truthful modulation.
This work thus aims at creating an artificial synapse that emulates as closely as possible its biological counterpart and presents a platform capable of creating a closed-loop neuronal micro-circuit based on neuromorphic devices.
1. Van De Burgt, Y. et al. A non-volatile organic electrochemical device as a low-voltage artificial synapse for neuromorphic computing. Nature Materials vol. 16 414–418 (2017).
2. Keene, S. T. et al. A biohybrid synapse with neurotransmitter-mediated plasticity. Nat. Mater. 19, 969–973 (2020).

The realization of bioinspired circuits that mimic the signal processing capabilities of the brain demands the reproduction of both short- and long-term aspects of synaptic plasticity on a single device level. The emergence of organic electronics has made available a host of new materials and devices with properties that are well-suited to address this challenge. We show how mixed conductivity in organic semiconductors can be leveraged to control neuromorphic functions similar to the basic short-term and long-term synaptic plasticity. Namely, we demonstrate that the timescales for functions such as depression, adaptation and dynamic filtering can be controlled by tuning the ratio of electronic/ionic mobility. We further show that 3D printing can be leveraged to develop new architectures for neuromorphic computing.

Materials processing and synthetic design serve as ideal avenues to control the transport properties of polymer-based mixed ionic/electronic conductors. Small changes in chemistry can affect the materials electronic mobility, swelling, ion uptake and stability. Devices based on these materials have thus opened up new opportunities in bioelectronics, energy, and neuromorphic computing. Associative learning, a critical learning principle to improve an individual’s adaptability, has been emulated by few organic electrochemical devices. However, complicated bias schemes, high write voltages, as well as process irreversibility hinder the further development of associative learning circuits. Here, we present a non-volatile organic electrochemical transistor (OECT) based on a poly(3,4-ethylenedioxythiophene):tosylate (PEDOT:Tos)/ Polytetrahydrofuran (PTHF) composite. This device can continuously and reversibly change its conductance state at a write bias less than 0.8 V and the state retention time can be longer than 200 min without decoupling the write and read operation. By incorporating a pressure sensor and a photoresistor into the gate terminal of volatile and non-volatile OECTs, a neuromorphic circuit is demonstrated with the ability to associate two physical inputs (light and pressure), which may have implications for biomimetic devices like electronics-skin and neuroprosthetics. To unravel the non-volatility of this material, UV-Vis-NIR spectroscopy, X-ray photoelectron spectroscopy (XPS) and grazing-incidence wide-angle X-ray scattering (GIWAXS) are used to characterize the oxidation level variation, compositional change and the structural modulation of the PEDOT:Tos/PTHF films in various conductance states. These studies support the finding that non-crystalline PTHF retains injected cations, and thus allowing for the non-volatile memory behavior of the composite OECT devices. The implementation of the associative learning circuit as well as the understanding of the non-volatile material represent critical advances for organic electrochemical devices in neuromorphic applications.

Artificial intelligence applications have demonstrated their enormous potential for complex processing over the last decade. However, they are mainly based on digital operating principles while being part of an analogue world. Moreover, they still lack the efficiency and computing capacity of biological systems. Neuromorphic electronics emulate the analogue information processing of biological nervous systems.¹ Organic artificial synapses exhibit volatile as well as tunable/non-volatile conductance states that replicate the behavior of their biological counterparts.^2-6
In this work we present a path planning robot that uses a small-scale, locally-trained organic neuromorphic circuit to navigate through a maze. The neuromorphic circuit responds and adapts to environmental stimuli directly, as it is integrated with the robot sensors. The on-chip integration of sensor signals together with low-power organic neuromorphic electronics paves the way toward stand-alone, brain-inspired computing circuitry in autonomous and intelligent robotics.⁷


References
C. Mead, Neuromorphic electronic systems. Proc. IEEE 78, 1629-1636 (1990).
P. Gkoupidenis et al. Synaptic plasticity functions in an organic electrochemical transistor. Appl. Phys. Lett. 107 (26), 263302 (2015).
P. Gkoupidenis et al. Neuropmorphic functions in PEDOT:PSS organic electrochemical transistors, Adv. Mater. 27 (44), 7176 (2015).
Y. van de Burgt et al. A non-volatile organic electrochemical device as a low-voltage artificial synapse for neuromorphic computing. Nat. Mater. 16, 414 (2017).
E. J. Fuller et al. Parallel programming of an ionic floating-gate memory array for scalable neuromorphic computing, Science 364 (6440), 570 (2019).
A. Melianas et al. Temperature-resilient solid-state organic artificial synapses for neuromorphic computing, Sci. Adv. 6, (27) eabb2958 (2020).
C. Wan et al. Artificial Sensory Memory. Adv. Mater. 1902434, 1–22 (2019).

Traditional computing systems are unable to capture the capability of the brain in real world information processing as evidenced by the anticipated end to Moore’s law. Neuromorphic computing aims to mimic the efficiency of the brain in complex classification tasks by applying learning directly in hardware and is promising for local edge computing applications such as smart point-of-care devices (1). However, straightforward tuneability, low power consumption and high accuracy are necessary for successful implementation of such applications (2). Recently, organic materials have emerged as building blocks for neuromorphic systems due to their low operation voltage, biocompatibility and excellent tunability properties (3). Nevertheless, fully autonomous bioelectronic applications demand not only the acquisition of biological signals, but also local data processing, storage and the extraction of specific features of merit (4). Here we show a smart biosensor based on organic neuromorphic devices that can be fully trained in hardware to classify a model disease (e.g. cystic fibrosis). We use ion selective electrodes that are able to detect physiological levels of potassium and chloride and serve as the input to the hardware-implemented neural network. We demonstrate the flexibility and versatility of the neuromorphic system by on-chip retraining with different input signals as well as by classification of logic gates. These results pave the way for adaptive biosensors, smart point-of-care devices and personalized healthcare.


1. Fuller et al., Science, 2019
2. Van de Burgt et al., Nature Materials, 2017
3. Van de Burgt et al., Nature Electronics, 2018
4. Van Doremaele et al., J. Mater. Chem. C, 2019

Silicon-based microstructures are gaining importance in fundamental neuroscience and biomedical research. Precise and long-lasting neuro-electronic hybrid systems are at the center of research and development in this field. Nowadays, the best approach to study the electrophysiological activity of neurons in vitro and in vivo is based on planar microelectrode arrays (MEA) or field-effect transistors (FET) which can be integrated with microfluidic devices. However, the weak coupling between cell membrane and electrode surface is one of the major limiting factors and technology of 3D nanostructures for cell-chip coupling is currently a vivid field of investigation. Our present study focuses on the investigation of cell-chip interfaces with 3D nanoelectrodes for extracellular recordings. In particular 3D nanostructures with a high aspect ratio have become a suitable interface between neurons and electronic devices, since the excellent mechanical coupling to the neuronal cell membrane allows very high signal-to-noise ratios. In the light of an increasing demand for a stable, non-invasive and long-term recording at subthreshold resolution, we present a combination of vertical nanostraws with nanocavities. These structures provide a tight coupling with rat cortical neurons, resulting in high amplitude sensitivity and post-synaptic resolution capability, as directly confirmed by combined patch-clamp and MEA measurements.

Organisms display a remarkable ability to perform complex functions in unstructured settings thanks to their co-evolved neuromechanical systems. Filtering, thresholding, and nonlinear amplification are common features found in organisms' mechanosensing. In contrast to engineering systems that rely on electrical filtering, organisms often employ distributed mechanical filters intimately integrated within their material architecture to extract only the necessary information from the environment. Filtering and signal thresholding are often achieved by careful positioning of geometrical features in the form of hairs or cilia spiders' legs, bats' wings, and fish's skins to sense the surrounding conditions mechanically.
An equally important aspect in neuromechanical systems is the capacity to form and retain memory that classifies different external stimuli. A crucial such mechanism is the ability to recall events from partial information, often described as associative memory. Integrating mechanosensing and memory formation as intrinsic properties in soft systems constitutes a new frontier in the development of engineered materials with potential applications in robotics, morphing structures, and wearables.
We report on multistable metamaterial architectures exhibiting hybrid units that enable filtering, thresholding, transduction, and associative memory formation, providing a route to synthetic neuromechanical systems. Our hybrid constitutive units collocate a geometrically bistable structure, a piezoresistive sensor, and a polymeric memristor. Specifically, we leverage the local bistability of constitutive units to filter external inputs, which are only transduced once the specific force threshold required to induce unit inversion (snap-through) is reached. The resulting unit inversion causes a nonlinear strain field that amplifies the external input, subsequently inducing transduction into an electrical signal leveraging collocated piezoresistive strips. The transduced signals trigger memristance changes which allow for memory retention. Notably, successive transduced signals accumulate changes of memristance, thereby enabling the formation of associative memories following a Hebbian rule within the paradigm of Hopfield networks. We demonstrate that the final accumulated signal on our memristors can be used to recall several spatiotemporal forcing patterns acting on metasheet featuring N hybrid units. Our metamaterials offer the rudiments of neuromechanical systems while requiring minimal external power and digital electronics, opening a route to engineered structures with organism-inspired functionality.

Spiking Neural Networks (SNN), provide distributed computational capability and emulate the brain processing principles. Compared to conventional Artificial Neural Networks (ANN) that encode the information in the magnitude of the signal, SNN function based on modulation of spike frequency and pulse width [1]. Due to intrinsic compatibility with brain signaling, one of the target applications of SNN hardware implementation, known as neuromorphic systems [2], are Brain-Computer Interfaces (BCI) [3]. With a handful of exceptions [4], such implementations are typically implemented using silicon technology, which is physically hard and rigid, and non-biocompatible, resulting in a permanent damage to the brain's soft tissue. Organic electronics are physically soft and flexible, biocompatible, and offer low-cost, large-area fabrication, for direct interface with biological tissues [5].

Synapses play a crucial role in neural networks by interfacing between adjacent neurons [1]. They are responsible for tuning the neural signals and, as such, are critical in memory, learning, and cognition.

We report on the fabrication of a physically flexible organic electronics, spiking synaptic circuits. Organic Field-Effect Transistors (OFETs) were fabricated on rigid glass and physically flexible 50 μm thin Polyimide substrates. OFETs’ architecture was bottom-gate top-contact, and fabricated through thermal deposition of Cr (3 nm)/Au (50 nm) as the gate electrode, 400 nm of Parylene diX-SR as the dielectric, 30 nm of dinaphtho[2,3b:2’,3’-f]thieno[3,2-b]thiophene (DNTT) as the p-type and N,N’-bis(n-octyl)-x:y,dicyanoperylene3,4:9,10-bis(dicarboximide) (PDI8-CN2, also referred to as N1200) as the n-type organic semiconductors. A 50 nm layer of Au served as the source and the drain electrodes. A layer of Parylene encapsulated the structure. Non-fabrication-integrated commercial capacitors were used to accelerate the fabrication and the characterization process. Future work will include all integrated, all-organic circuits.

The results demonstrate that the fabricated log-domain integrator synapse [6], a variant of the linear charge-and-discharge synapse, can implement a true linear integrator circuit. The circuit exploits the logarithmic relationship between subthreshold OFET gate-to-source voltages and their channel currents. It provides proportional output current based on continuously tunable synaptic weights. We also demonstrate the effects of synaptic weighting voltage, synaptic capacitance, and disparity in pre-synaptic signals on circuit time constant.
Crucially, we show that our organic synaptic circuits can provide time constants that are more than 200 times longer compared to silicon-based synaptic circuits, recorded in excess of 6 seconds. Also, theoretical calculations are commensurate with empirically recorded time constants.

We believe that this work is the first demonstration of a spiking organic electronic synaptic circuit. Time constants in the range of seconds are considered critical for future integration of synaptic neuromorphic circuits with biological neurons, for future implantable BMIs. Furthermore, with our previously demonstrated spiking organic electronic Axon-Hillock somatic circuits [7], it paves the way for future all-organic, flexible and biocompatible neuromorphic computing.

[1] F. Rieke and D. Warland, MIT Press, 1999.
[2] R.A. Nawrocki, et al., IEEE Trans. Ele. Dev.,63 (10), 2016.
[3] F. Boi, et al., Front. Neuro., 10 (563), 2016.
[4] R.A. Nawrocki, et al., IEEE Trans. Ele. Dev., 61 (10), 2014.
[5] G. Schwartz, et al., Nat. Comm., 4 (1), 2013.
[6] P. Merolla and K. Boahen, MIT Press, 2004.
[7] Hosseini, MJM, et al., J.Phys.D: Appl.Phys., 54, 104004, 2021.

Neuromorphic devices and circuits offer huge potential for hardware-based neuromorphic computing and next-generation, power-efficient artificial intelligence (AI). Organic electrochemical transistors (OECTs) and related devices emerged as an interesting class of devices for neuromorphic computing as the interplay between ions and electrons in OECTs can be considered biomimetic. The research on OECTs mainly focuses on developing powerful and versatile artificial synapses for conventional AI paradigms, e.g., feed-forward neural networks. However, because of their biomimetic operation, OECTs might also be used to build bio-inspired spiking neurons for next-generation power-efficient computing. Here, we show an OECT with photo-patternable solid electrolyte advancing towards the integration of bio-inspired neurons. We discuss how the neuromorphic properties of OECTs, i.e. the time constants of the ionic and electronic subcircuit of the device, can be controlled by process and device parameters. We also demonstrate the first step towards the integration of Morris-Lecar inspired neurons, which offer type-I and type-II excitability in their spiking behavior. In particular, we discuss how the specific properties of the solid-state electrolyte OECTs, such as threshold voltage, subthreshold slope, hysteresis, and off-current leakage, influence the function of the subcircuit of our biomimetic neuron. The solid-state electrolyte OECT paves the foundation towards the integration of neuromorphic devices beyond artificial synapses, and the small-scale integration of OECTs exhibit potential for building blocks of next-generation spiking neuron circuits.

Strongly correlated perovskite nickelates (RNiO₃) have attracted great interest in condensed matter sciences due to its unique electrical properties such as metal-to-insulator transition. Hydrogenation induced electron doping leads to a phase transition in this material system independent of temperature. The electrical conductivity of RNiO₃ is extremely sensitive to the electron occupancy of Ni 3d orbitals. The electron from hydrogen occupies Ni eg orbital and results in remarkable resistance change by several orders of magnitude. The proton resides on interstitial sites of NiO₆ octahedra and can be mobile under electric fields. In this work, we applied high-speed voltage pulses and measured the resistance changes of hydrogen doped NdNiO₃ devices fabricated with conducting oxide (FTO) electrodes. We will discuss the effect of voltage pulses on proton migration in the nickelate device, which enables thousands of analog states with highly tunable electronic structures. Interestingly, non-volatile memory trees can be demonstrated in perovskite nickelate device at room temperature by applying consecutive positive or negative voltage pulses emulating synaptic weight branching noted in animal brains. This behavior is akin to magnetic frustration in spin-glasses at low temperatures. The complex synaptic behavior in perovskite nickelate devices can open up new avenues for design of memories for neuromorphic computing.

In today’s computing architectures, fundamental physical limits are increasingly prohibiting performance improvements. At the same time, the amount of data, recorded by an ever rising number of sensors in various applications, is growing rapidly. In order for computers to keep up with this contradiction, new computing paradigms are required. Neuromorphic technologies are seen as key solutions as they can be employed at the edge of the sensors, rendering the need for storing the data or sending them to remote clouds. Augmenting these devices with emerging memory technologies empowers them with non-volatile and adaptive properties, which are desirable for low power and online learning requirements at the edge, independently from remote cloud computers. However, an energy- and area-efficient realization of such systems requires disruptive hardware and algorithm changes. Solutions based on memristive devices are in the focus of research and industry due to their low-power and high-density online learning potential. Specifically, the filamentary-type valence change mechanism (VCM) memories have shown to be promising candidates. Because of CMOS compatibility, HfO₂-based cells are among the most researched devices and are nearing industrial maturity. In consequence, physical models capturing a broad spectrum of experimentally observed features such as the pronounced cycle-to-cycle and device-to-device variability are required for accurate evaluation of the proposed concepts.
In this study, we present an in-depth experimental analysis of device-to-device and cycle-to-cycle variability of filamentary-type bipolar switching HfO₂/TiO_x nano-sized crossbar devices. A match of the experimentally obtained statistics and variabilities to our physically motivated, variability-aware JART VCM compact model [1] is established. The high degree of agreement between model and experiment allows us to evaluate the concept of parallel memristive devices forming a single synapse. Both experimental and theoretical results are presented and conditions for favorable properties of such synapses are defined. The interplay of cycle-to-cycle and device-to-device variability is analyzed in detail and descriptive figures of merit are defined, since the proposed concept should retain its functionality in case of devices with lower variability.
These parallel compound synapses form a synaptic array which is at the core of many neuromorphic chips. We exploit the cycle-to-cycle variability of the single devices for stochastic online learning which has shown to increase the effective bit precision of the synapses [2]. Finally, we demonstrate the stochastic switching operation as synapse update method in a Spiking Neural Network which classifies two patterns. In this small scale demonstrator, perfect accuracy is achieved, proving the feasibility of this alternative synapse concept. [3]

[1] C. Bengel, A. Siemon, F. Cüppers, S. Hoffmann-Eifert, A. Hardtdegen, M. von Witzleben, L. Hellmich, R. Waser and S. Menzel, Variability-Aware Modeling of Filamentary Oxide based Bipolar Resistive Switching Cells Using SPICE Level Compact Models, TCAS 1 67, 4618-4630 (2020)
[2] S. Gupta, A. Agrawal, K. Gopalakrishnan and P. Narayanan, Deep Learning with Limited Numerical Precision, 1737–1746 (2015)
[3] C. Bengel, F. Cüppers, M. Payvand, R. Dittmann, R. Waser, S. Hoffmann-Eifert and S. Menzel, Utilizing the Switching Stochasticity of HfO2/TiOx-Based ReRAM Devices and the Concept of Multiple Devices for the Classification of Overlapping and Noisy Patterns, Frontiers in Neuroscience 15, 621 (2021)

Next generation of energy storage may largely benefit from fast Li⁺ ceramic electrolyte conductors to allow for safe and efficient batteries. With recent discoveries in thin film processing solid-state lithium ion conductors, such as Li-garnets and LIPON or LiSICON-based solids, have been recently considered as candidate materials not only for next-generation solid-state batteries but also for neuromorphic computing via memristors owing to the fast ionic transport in the solid-state electrolyte.
In the first part of this talk, we review current needs and challenges in the field of solid state batteries. We underline the advantages of various Li solid-state conductor materials and reflect on opportunities of thin film processing, being a requirement to define precisely the Lithium stoichiometries and related electronic state changes for transition metal ions, miniaturize the device, and reach high energy/information densities for energy storage, computation, and sensing.
In the second part, we focus on Li-film processing and controlling Lithium stoichiometries to reach fast conductive phases for Li garnets and Li titanates as solid state battery components, and memristive neuromorphic computing units. Insights on structure-phase-transport interaction and implications on performances will be exemplified for energy storage aiming high energy densities, and modulations of synaptic artificial weights through lithium induced metal-to-insulator transitions in lithium titanate memristors.

Artificial neural networks based on electrochemical random-access memory (EC-RAM) represent an emerging strategy for low power analog computing. EC-RAM synapses operate by virtue of linear/symmetric resistance tuning through electrochemically gated control of the charge state of a resistive channel. The exploration of promising active materials for EC-RAM functionality was previously limited to semiconducting organic polymers and metal oxides. Here we demonstrate a Ru-based Prussian Blue Analogue (RuPBA), an electroactive mixed valence semiconductor with a cyanide-bridged framework structure, as a new candidate material for EC-RAM devices. Our fabrication method employs inkjet printing to prepare both the RuPBA and ionogel electrolyte films. The RuPBA channel resistance depends on the electrochemically tunable ratio of Ru^II to Ru^III and the degree of cation intercalation in the micropores of the material. The relationship between charge state, resistance, and electronic structure is elucidated using in-operando Raman and UV/Vis/NIR spectroscopy. The RuPBA devices exhibit excellent linearity, symmetry, and consistency during pulsed voltage programming, as well as long term retention of conductance state.

Among the outstanding capabilities of biological brains is their ability to visualize, memorize, and recognize objects in erratic and unstable environments. The strength, versatility, and fault-tolerance of brains result from the sheer number of synaptic connections between neurons that is estimated to be ~10¹⁴-10¹⁵ in a human brain. Reactivation of specific groups of neurons and the strength of connections between them play important roles in memorization and information processing. In the age of new information and communications technologies, these traits, among others, serve as motivation for design choices in development of bio-inspired computing technologies. Neuromorphic computing and artificial neural networks today represent an example of the most successful bio-inspired technologies that excel in a broad range of classification and regression tasks.
A large portion of current efforts to reach the goal of neuromorphic systems focuses on the use of ordered nanodevice structures, most notably the memristor crossbar array. Lithographic fabrication of such ordered structures inherently limits neuromorphic systems' complexity and scale and cannot mimic the sheer scale of complex interconnections of biological neural networks. In this study, two-dimensional, stochastic, self-assembled, bio-inspired, highly interconnected silver sulfide (Ag₂S) nanowire networks are studied experimentally and through simulation to investigate their usefulness for neuromorphic hardware applications, potentially for use in Reservoir computers. Resistive switching behaviors resulting from the electromigration interaction between memristive nanowires are explored. Single memristive Ag₂S nanowire, as well as small, randomly assembled networks of NWs with countable numbers of components, are also studied. We further compare experimental measurements with predictions from a novel simulation model, the "stochastic drift R_ON memristor model". This computational model suggests that R_ON, the resistance observed in the memristors’ ON state, is not a constant but rather a function of the volume fraction of reduced silver ions in the Ag₂S matrix. In the experimental memristive systems described, several crucial synaptic functions are demonstrated, such as learning, forgetting, paired-pulse facilitation, and short-term plasticity. Some aspects of the physical mechanisms of the devices are also inferred from the observed behaviors. The self-assembled nanodevices and device networks described in this research are first steps toward three-dimensional randomly self-assembled nanocomposite memristor networks for neuromorphic computing applications. Such materials represent promising candidates for next-generation biomimetic hardware. A critical issue yet to be resolved involves interfacing such memristive device networks with multi-electrode arrays such that finer-grained substates of the networks’ electrical behaviors can be simultaneously and continuously read and routed for biologic-like computational tasks. This work aims to provide a convenient, low cost and low-power physical implementation for Reservoir Computing research with potential applications in communications, signal processing, and control.

As conventional silicon based computing techniques are reaching their limits in terms of energy consumption and density[1], neuromorphic platforms are emerging as new computational paradigm[2], with the ultimate goal of replicating the efficiency of the brain in terms of parallel computation. Among the plethora of the available materials and structures, organic electrochemical transistors (OECTs) have emerged for their natural ionic-to-electronic signal transduction, biocompatibility and possibility of operating in aqueous environment[3]. In particular, PEDOT:PSS based OECTs have been exploited to recapitulate neurotransmitter-mediated synaptic plasticity[4], in which the oxidation of a neurotransmitter modulates the doping state of the transistor, resulting in a change in device conductivity, de facto emulating the synaptic conditioning naturally occurring in biological synapses.
While successfully recapitulating spike-dependent plasticity, this approach still lacks the self-regulation mechanisms naturally found in neuronal networks: when the post-synaptic neuron is engaged in the neurotransmitter-mediated communication, receptor recycling occurs[5], while astrocytes provide a feedback onto the presynaptic terminal, to reduce or increase the release of such neurotransmitter[6]. Moreover, the PEDOT:PSS-based neuromorphic transistor, if used as a building block of an artificial network, would lead to a continuous de-doping of the devices, ultimately leading to the exhaustion of the network dynamics, because of the unidirectional modulation of the synaptic weight.
Taking inspiration from the nigrostriatal DA pathway, in which the endogenously generated H2O2 activates K+ channels, inhibiting DA neuron firing[7], a closed-loop synapse is proposed, in which the bidirectional modulation of the synaptic weight is shown. In particular, the output current of the device closes a feedback loop, allowing for a controller to drive a microfluidic system. As a result, an influx of dopamine (DA), and its subsequent oxidation, is used to cause the de-doping of both OECT gate and channel, while the inverse reaction can be induced by the introduction of hydrogen peroxide (H2O2).
The presence of the control system proposed here introduces the astrocytes-like effect of presynaptic input regulation, naturally found in neural networks. This enables for synaptic weight control while avoiding the uncontrolled doping/de-doping of the PEDOT:PSS.
Ultimately, this platform paves the way for the creation of biomimetic complex systems. A network of artificial neurons can be designed in which the output depends on the presence of a neurotransmitter while each node can still be individually controlled, to induce and study topology-dependent dynamics propagation across the network.

1. Hassan, S., Humaira & Asghar, M. Limitation of Silicon Based Computation and Future Prospects. in 2010 Second Int. Conf. Commun. Softw. Netw. 559–561 (2010). doi:10.1109/ICCSN.2010.81
2. van de Burgt, Y., Melianas, A., Keene, S. T., Malliaras, G. & Salleo, A. Organic electronics for neuromorphic computing. Nat. Electron. 1, 386–397 (2018).
3. Friedlein, J. T., McLeod, R. R. & Rivnay, J. Device physics of organic electrochemical transistors. Org. Electron. 63, 398–414 (2018).
4. Keene, S. T., Lubrano, C., Kazemzadeh, S., Melianas, A., Tuchman, Y., Polino, G., Scognamiglio, P., Cinà, L., Salleo, A., van de Burgt, Y. & Santoro, F. A biohybrid synapse with neurotransmitter-mediated plasticity. Nat. Mater. 19, 969–973 (2020).
5. Jung, N. & Haucke, V. Clathrin-Mediated Endocytosis at Synapses. Traffic 8, 1129–1136 (2007).
6. Newman, E. A. New roles for astrocytes: Regulation of synaptic transmission. Trends Neurosci. 26, 536–542 (2003).
7. Rice, M. E. H2O2: A Dynamic Neuromodulator. The Neuroscientist 17, 389–406 (2011).

Resistive switching (RS) cells are promising as synaptic devices owing to analog controllability of the cell resistance. However, most transition metal oxide-based RS cells often show an abrupt decrease in the cell resistance (set process), resulting in the difficulty of the analog control of the cell resistance by the applied voltage [1]. Moreover, although some studies have reported gradual set processes, the mechanism of the gradual set processes is not clear [2]. Therefore, no design rules exist to improve the analog RS characteristics. In our previous study, we reported that analog RS characteristics can be obtained by adjusting the applied voltage after the unique forming process (semi-forming) in Pt/TaO_x/Ta₂O₅/Pt cells [3]. In this study, we mainly investigated the impacts of oxygen composition of the TaO_x layer on the semi-forming characteristics for the purpose of clarifying the mechanism of the semi-forming and subsequent analog RS phenomena.
We fabricated platinum (Pt)/tantalum sub-oxide (TaO_x)/tantalum oxide (Ta₂O₅)/Pt stack cells on a SiO₂/Si substrate. A 60-nm-thick Pt bottom electrode (BE) was deposited by DC sputtering. Then, a 10-nm-thick Ta₂O₅ and a 20-nm-thick TaO_x layer were deposited by reactive RF sputtering with different oxygen gas flow rates. The oxygen composition of the TaO_x layer was controlled by appropriately adjusting the oxygen gas flow rate between 0.8 sccm and 1.2 sccm. Finally, 25-nm-thick Pt top electrodes (TE) with various diameters (100–300 μm) were formed by electron beam evaporation. We evaluated forming characteristics by performing a current sweep on the Pt/TaO_x/Ta₂O₅/Pt cells with the BE grounded at room temperature.
Two types of initial transition processes occur in the Pt/TaO_x/Ta₂O₅/Pt cells. One is the transition process to a low resistance state (LRS) less than 100 Ω, which is conventionally called a forming phenomenon. The other is the transition process to a higher resistance state than the LRS, which is defined as a semi-forming phenomenon. Next, we examined dependences of the oxygen composition (x), initial resistance (R_ini), and occurrence frequency of the semi-forming phenomenon (F_semi) on the oxygen flow rate during deposition of a TaO_x layer (f_Ox). First, x is proportional to f_Ox, and the f_Ox range of 0.8–1.2 sccm corresponds to the x range of 1.3–2.0. Second, as x linearly increases, R_ini exponentially increases and F_semi tends to decrease. Here, we have reported that R_ini is mainly dominated by the hopping conduction through oxygen vacancies (V_Os) in the Ta₂O₅ layer [4]. The density of localized states corresponding to the V_O density was estimated from the temperature dependence of R_ini in each Pt/TaO_x/Ta₂O₅/Pt cell, resulting in that the density of localized states exponentially decreased from roughly 10²² cm⁻³eV⁻¹ to 10¹⁹ cm⁻³eV⁻¹ as x linearly increased from 1.3 to 1.7. In other words, the density of V_Os supplied from the TaO_x layer to the Ta₂O₅ layer decreases with increasing x.
From the above results, although the semi-forming phenomenon is more likely to occur as more V_Os are supplied to the Ta₂O₅ layer, the more supplied V_O causes the decrease in R_ini, bringing about a decrease in the resistance range of RS operation. To obtain the semi-forming and subsequent analog RS characteristics in the high-resistance range for reduction of power consumption, it is necessary to appropriately control the oxygen composition of the TaO_x layer.

[1] D. Ielmini, Microelectronic Engineering, 190, 44 (2018).
[2] R. Waser et al., Faraday Discuss. 213, 11 (2019).
[3] T. Miyatani et al., 2020 virtual MRS spring/fall meeting, F. NM07. 06. 06.
[4] T. Miyatani et al. Jpn. J. Appl. Phys. 58, 090914 (2019).

Phase change memory (PCM) is a promising candidate for non-von Neumann based analog in-memory computing – particularly for inference of previously-trained Deep Neural Networks [1,2], as well as Neuromorphic computing [3,4]. Many factors including resistance values, memory window, resistance drift, read noise, and programming accuracy impact the performance of PCM in analog neuromorphic computing applications. We previously showed that resistance drift and noise of a mushroom-type PCM device can be reduced by introducing an additional projection liner [5]. The projection liner comprised of a non-phase change material could decouple the read operation from the write operation and help mitigate the non-ideal attributes such as drift and noise.

Here, we perform a systematic study of the electrical properties of these PCM devices—including resistance values, drift, noise, and endurance—and discuss the implications for in-memory computing. We optimize the PCM and projection liner materials to have effective drift mitigation while retaining a large enough memory window. We also examine how endurance cycling affects the phase change materials and the electrical characteristics of the devices. The results show that the liner is helping to maintain a rather constant reset resistance throughout the endurance cycles. With a widened memory window, no degradation in the drift mitigation through the cycling, and reduced read noise, these cycled PCM devices are shown to deliver excellent balance of performance for analog AI applications. Finally, the performance of these PCM devices for neural network inference is analyzed for state-of-the-art image processing and natural language processing tasks.

[1] P. Narayanan, et al., Fully On-Chip MAC at 14nm Enabled by Accurate Row-Wise Programming of PCM-Based Weights and Parallel Vector-Transport in Duration-Format, Proc. VLSI Tech. Symposium, 2021
[2] R. Khaddam-Aljameh, et al., HERMES Core – A 14nm CMOS and PCM-based In-Memory Compute Core using an array of 300ps/LSB Linearized CCO-based ADCs and local digital processing, Proc. VLSI Circuits Symposium, 2021
[3] I. Boybat, et al., Neuromorphic Computing with Multi-Memristive Synapses, Nature Communications 9, 2514, 2018
[4] G. W. Burr, et al., Neuromorphic computing using non-volatile memory, Advances in Physics: X, 2:1, 89, 2017
[5] R. Bruce, et al., Mushroom-type phase change memory with projection liner: an array-level demonstration of conductance drift and noise mitigation, Proc. IRPS, 2021

For numerous Radio-Frequency applications such as medicine, RF fingerprinting or radar classification, it is important to be able to apply Artificial Neural Network on RF signals. In this work we show that it is possible to apply directly Multiply-And-Accumulate operations on RF signals without digitalization, thanks to Magnetic Tunnel Junctions (MTJs). These devices are similar to the magnetic memories already industrialized and compatible with CMOS.
These devices are promising RF emitters. Like biological neurons, MTJs are non-linear oscillators, and it was already proven that they can efficiently emulate neurons in the framework of reservoir computing. These junctions can also be used as detectors and rectify RF signals.
We show experimentally that a chain of these MTJs between 250 and 400 nm of diameters can rectify simultaneously different RF signals, and that the synaptic weight encoded by each junction can be tune with their resonance frequency. Our architecture benefits both from frequency-multiplexing of the RF inputs and from the frequency selectivity of MTJs to route the information with a compact architecture.
Through simulations we train a layer of these junctions to solve the dataset “digits” with an accuracy as good as software neural networks (99.96 %). Finally, we show that our system can scale to multi-layer neural networks using MTJs to emulate neurons. In the future these MTJs could be scaled down to 10 nm.
Our proposition is a parallel and compact system that allows to receive and process RF signals in situ and at the nanoscale.
This work was supported by the European Research Council ERC under Grant bioSPINspired 682955, the French ANR project SPIN-IA (ANR-18-ASTR-0015) and the French Ministry of Defense (DGA).

Memristor technology provides a more efficient solution to emulating the behaviors of synapses and neurons in neuromorphic systems compared to conventional CMOS technology. Memristive resistances can store synaptic weights, while memristive nonlinearity can be exploited for neuron circuits. Achieving the advanced learning capabilities and adaptability of the brain requires further circuitry to leverage the characteristics of these memristive devices. Basic spike-based learning algorithms such as spike-timing-dependent plasticity (STDP) have previously been demonstrated. In addition, circuits for exploiting novel device dynamics for flexible and powerful training rules will promote computationally efficient intelligence that approaches that of the human brain.
Spike-based encoding methods enable more efficient data processing when applied to memristor-based hardware for artificial deep (DNNs) and spiking neural networks (SNNs). In these designs, the memristor-based synapses take in binary spike trains rather than analog voltages, avoiding power-hungry analog-to-digital conversions and thus elevating energy efficiency. Encoding data as a binary spike train has historically involved generating proportional spike rates. However, an increase in data value corresponds to an increase in the number of spikes. Temporal coding encodes data as discrete timings of voltage spikes for a much more efficient and precise method of representing data. For example, time-to-first-spike (TTFS) encodes data as the arrival time of the first spike during a predefined time period [1]. It minimizes spike numbers by using only one spike to carry data.
In realizing such neuromorphic systems, new ideas and approaches must be advocated. Architectures consisting of simple crossbar synapses and integrate-and-fire neurons should be adapted in favor of more biologically derived but often oversimplified models. Each component of a neural system incorporates its own unique dynamics and offers opportunities for a behaviorally parallel circuit to be developed and incorporated into a sophisticated neuromorphic design. Furthermore, the fast-shrinking factor size of novel devices promotes the potential for computationally dense neuromorphic building blocks. Following these methodologies, we proposed a memristor-based single-spiking processing engine with a set of circuitry designs for supporting DNN inference with the TTFS scheme, leading to faster computations and lower energy consumption [2]. Compared to conventional memristor-based dot product engines that encode data as the amplitude of voltages or currents, this design leverages the advanced spike-based encoding to shorten the computational period and pursue high efficiency. Furthermore, we demonstrated in-situ learning in a memristor crossbar using a methodology for single-spike-encoded STDP that could be incorporated into a TTFS system [3].
In our future exploration, we will investigate techniques for building large-scale memristor-based neuromorphic systems with online learning capabilities. The targeted system should support both the inference of input data and the update of neural network parameters. Circuit/architecture-level designs are required to implement either supervised or unsupervised learning rules with memristors while ensuring the forward and backward data paths do not interfere with each other.
References:
[1] S. Thorpe, A. Delorme, and R. Van Rullen, “Spike-based strategies for rapid processing,” Neural Networks, vol. 14, no. 6-7, pp. 715-725, 2001.
[2] Z. Li, B. Yan, and H. Li “ReSiPE: ReRAM-based Single-Spiking Processing-In-Memory Engine,” in 2020 57^th ACM/IEEE Design Automation Conference (DAC), pp. 1-6, 2020.
[3] B. Taylor, A. Shrestha, Q. Qiu, and H. Li. “1S1R-Based Stable Learning Through Single-Spike-Encoded Spike-Timing-Dependent Plasticity.” 2021 IEEE International Symposium on Circuits and Systems (ISCAS), pp. 1-5, 2021.

Artificial intelligence (AI) today outperforms humans for a variety of applications, such as computer vision and natural language processing. Higher accuracy comes at the cost of increased computational complexity and model size, stressing already limited memory access.
RRAM-based in-memory computing (IMC) provides a dense and parallel structure to achieve high performance and energy efficiency in AI computing. The increased energy-efficiency is attributed to a full-custom design following the assumption that all weights are stored on-chip. However, such an assumption results in a significant area overhead, especially when the AI model size is rapidly increasing. Hence, model compression (e.g. pruning and quantization) is necessary for RRAM-based in-memory acceleration.
When an AI model is mapped to realistic RRAM devices, it will suffer from various non-idealities, such as statistical variations, quantization error, stuck-at-faults, and limited resistance ratio. These issues pose a significant challenge to designing robust RRAM-based IMC architectures. For instance, statistical variations in RRAM cause deviations in the programmed resistance, leading to a significant loss in post-mapping accuracy (i.e., accuracy in the presence of RRAM variations) for AI models. To mitigate the post-mapping accuracy loss, variation-aware training (VAT) is employed. VAT exploits the inherent redundancy in the AI model by embedding the device variations, based on a log-normal or normal distribution model, into the training process to achieve a variation-tolerant model, with no need of re-training for each individual RRAM chip.
Under model pruning and quantization, the efficacy of conventional VAT methods is reduced due to increased model sensitivity to variations, although a lower precision may help improve the tolerance. These algorithmic and device issues interact with each, affecting the optimal design choice of RRAM-based IMC. Hence, there is a need for a more systematic solution for robust RRAM-based in-memory computing with sparse and quantized AI models.
In this work, we propose a model stability-based VAT approach to recover the post-mapping accuracy with model compression. The proposed method utilizes VAT to train the compressed model with different scales of device variations and search for the most stable model to improve post- mapping accuracy. For a given model, higher model stability implies better tolerance of variations and thus, higher post-mapping accuracy. The model stability is visualized by the loss landscape and evaluated by the roughness score. A lower roughness score indicates a smoother loss landscape and a more stable model. We utilize a structured pruning method and model quantization to compress the AI model. The pruning method considers the mapping of the model on to the RRAM crossbar for best IMC performance. We illustrate that pruning results in a less stable model, while quantization improves the model stability.
For accurate modeling, we measured statistical variations from a 65nm CMOS/RRAM 1T1R chip and developed a cross-layer benchmarking tool. This tool incorporates RRAM device models, IMC architecture, pruning and quantization, and evaluation of model stability and outputs post-mapping accuracy, hardware performance, and model stability.
The major contributions of this work are as follows: (1) Based on the roughness score and loss landscape, we present that quantization (low-precision) improves the model stability, but pruning reduces the stability; (2) We propose a model stability-based VAT method, which searches the most stable model under variations to achieve the best post-mapping accuracy, without knowing the exact amount of RRAM testing variations upfront; and (3) We demonstrate that the proposed method achieves up to 19%, 21%, and 11% improvement in post-mapping accuracy on CIFAR-10, CIFAR-100, and SVHN datasets, respectively.

A great challenge in the field of neurocomputing is to mimic the brain behaviour by implementing artificial synapses and neurons directly in hardware. So far, many works were devoted to the realisation of artificial synapses but the realisation of artificial neurons in hardware is still scarce. This work shows that Mott insulators are interesting inorganic materials to consider in order to achieve this goal.
Mott Insulators experience resistive switching above a threshold electric field [1-4]. This resistive switching is related to the creation of hot carriers leading to an electronic avalanche breakdown and a burst of current in the electronic circuit (Firing event) [2]. Moreover, under a train of electric field pulses, narrow gap Mott insulators behave as a leaky integrators of hot carriers and this phenomenon could be modelled assuming the creation of a metallic phase at the local scale [3,4]. Narrow-gap Mott insulators subjected to electric pulsing can therefore implement the three basic functions Leaky Integrate and Fire (LIF) of artificial neurons [5]. A simple two terminal device made of Mott insulator thin film can therefore be considered as a promising LIF artificial neuron. This work shows the realisation and tests of such downscalable artificial neurons made with Mott insulators. [6]
References:
[1] E. Janod, et al., Adv. Funct. Mater. 2015, 25, 6287.
[2] V. Guiot, et al. Nat Commun 2013, 4, 1722; P. Diener, et al., Phys. Rev. Lett. 2018, 121, 016601; D. Babich, et al. , Physics and Simulation of Optoelectronic Devices XXVIII; 2020, 112740B..
[3] P. Stoliar et al., Adv. Mater. 2013, 25, 3222.
[4] P. Stoliar et al Phys. Rev. B 2014, 90, 045146.
[5] P. Stoliar et al., Adv. Funct. Mater. 2017, 27, 1604740; F. Tesler, et al., Phys. Rev. Appl. 2018, 10 (5), 054001; C. Adda et al. J. Appl. Phys. 2018, 124 (15), 152124.
[6] C. Adda et al., MRS Com. 2018, 8, 835–841.

In recent years, the improvement of traditional computing machines has been assisted by the Moore’s law. However, nowadays, further shrinking of device dimensions is currently impeded by the reaching of physical limits. Additionally, the separation between storage and processing of data in von Neumann classical machines is power consuming, in contrast with the trend of reduction of energy intake for greener electronics. The human brain is an example of how sophisticated tasks may be accomplished by a computing engine with reduced power consumption. Then, neuromorphic computing can provide new precious insights, ready for being hardware-translated. In this respect, oscillatory neural networks (ONNs) are appealing ultra-low-power neuromorphic architectures able to perform intelligent tasks, such as pattern recognition [A]. In the present contribution, we propose the innovative concept of implementing ONN by Beyond Complementary-Metal-Oxide-Semiconductor (CMOS) devices. In particular, we show experimental and simulation results of oscillators based on vanadium dioxide (VO2). VO2 is a transition metal oxide (TMO) that undergoes a phase transition from a high-resistive monoclinic (M1) crystal structure to a low-resistive tetragonal rutile-like (R) one. This phase transition takes the name of insulator-to-metal transition (IMT) and it is triggered by temperature. The IMT brings a resistive switching in VO2, where the resistivity change may be up to several orders of magnitude. It has been observed that the switch in resistance occurs as well in VO2 two-terminal devices, where self-heating is thought to play the main role. This occurrence is a key factor in implementing highly compact and scalable oscillators. We will investigate devices fabricated on Si-compatible technology on SiO2 substrate [B]. The lattice mismatch between the VO2 layer and the SiO2 substrate gives rise to a typical granular texture. We will use dedicated technology computer-aided design (TCAD) 3D electrothermal simulations of VO2 devices [C] and of the associated VO2 oscillators to assess the impact of this grainy structure over the device performances with a multi-scale approach that considers material properties and device architectures. Our findings will help shed light on the entangled thermal and electrical behavior of VO2 oscillators and can help provide guidelines for the successful implementation of ONN technology.
[A] A. Todri-Sanial, S. Carapezzi, C. Delacour, M. Abernot, T. Gil, E. Corti, et al., “How Frequency Injection Locking Can Train Oscillatory Neural Networks to Compute in Phase,” preprint, hal-lirmm.ccsd.cnrs.fr/lirmm-03164135, 2021.
[B] E. Corti, J. A. Cornejo Jimenez, K. M. Niang, J. Robertson, K. E. Moselund, B. Gotsmann, et al., “Coupled VO2 Oscillators Circuit as Analog First Layer Filter in Convolutional Neural Networks,” Front. Neurosci., vol. 15, 2021
[C] S. Carapezzi, C. Delacour, G. Boschetto, E. Corti, M. Abernot, A. Nejim, et al., “Multi-Scale Modeling and Simulation Flow for Oscillatory Neural Networks for Edge Computing,” 19th IEEE Interregional NEWCAS Conference, 2021.

Biological nerves can efficiently process complex information in real-time. To implement biological event-driven and parallel operation artificially, organic nervetronics have been upsurged. The artificial synapses can be combined with sensors and actuators to emulate the functions of biological sensory and motor neurons with a simple circuit structure. Furthermore, compared to CMOS-based von Neumann computing systems, they consume 10^-6 times less energy to operate. This organic nervetronics can be a new strategy for soft robotics and neuroprosthetics by emulating biological neuroplasticity events that have not been existed nerve before.
Here, we developed an artificial nerve that emulates biological afferent and efferent nerve. The flexible artificial afferent nerve consists of a resistive pressure sensor, organic ring oscillator, and artificial synapse. Sensory signals are converted into spike signals and processed in the artificial synapse, thus movement, simultaneous pressure, and braille characters can be detected through an artificial afferent nerve. The connection of biological organs and artificial afferent nerves shows a hybrid reflex arc that shows future applicability for neural prostheses. The movement of the detached cockroach leg can be controlled by external sensory information. Also, an artificial auditory system was developed by integrating a triboelectric nanogenerator an artificial synapse. The optoelectronic artificial efferent nerve can be constructed by integration of a photodetector, stretchable artificial synapse, and a polymer actuator (role of an artificial muscle). This optoelectronic artificial sensorimotor nerve emulates optogenetically engineered neurons and synapses. The contraction of artificial muscles well imitates biological muscle contraction. Furthermore, the system can distinguish alphabet Morse code for optical wireless communication method. Our flexible organic artificial nervous systems suggest a promising strategy for bioinspired electronics, soft robotics, and neuroprosthetics. Neuroplasticity-based artificial nerves can be used for restoring the biological afferent and efferent nerve response which can pave a new way to technology without using bulky electronic devices.

The impact of stress on the properties of thin film materials plays a crucial role in the performance of integrated devices. Indeed, in the last decades, stress engineering has been purposefully used to enhance the conduction of CMOS transistors thanks to increased carrier mobility under stress [1, 2]. Drawing inspiration from this successful use of stress in integrated devices, we turn our focus to integrated analog synapses used in Artificial Neural Networks which are key to bridging the computing bottleneck in Von-Neumann architectures [3]. In particular, the stress-based control of memristors is of interest. Not only is static stress targeted, but an external pressure source e.g. from an actuator or from a sound wave can be used as an additional input to the individual synapses in the learning process [4].
In this work we investigate the change in electrical conductivity caused by varying the pressure applied to integrated Ferroelectric Memristor devices. In these devices fabricated on Si with a CMOS compatible process, the switching of the polarization of the 3.5 nm Hf_xZr_1-xO₂ ferroelectric thin film layer in the stack has an influence on the electrical conduction. The state of the memristor is written by applying an electric field and it is read out with the measured current [5].
For this study, we show the development of a custom indenter setup which uses a flat sapphire tip 100 um in diameter to apply pressure (forces potentially up to 20 N). A calibrated load cell and motorized stage allow for a PID-controlled tuning of the applied force and thus exerted pressure on the device under test.
Increase of the Low Resistive State (2 MOhm LRS) up to 10% per 0.7 N applied force (0.56 GPa) on devices of 40 um diameter can be shown. On the other hand, the High Resistive State (HRS) is much less impacted by the pressure. By releasing the pressure and electrically cycling the device, the change in resistive state is reversible and the device operation is preserved after the pressure cycling. Changes up to 60% can be measured, however at the cost of a non-reversible state behavior or loss of the device after too much cycling. Variation in device dimensions and application of stress is used to investigate the physical mechanism behind the change in resistance.
Moreover, we propose a novel device concept where integrated piezoelectric actuators at the low micrometer and nanometer scale can be used to apply stress on the Ferroelectric Memristors [6]. When used as a synapse, the 2-port device gains through the actuator an additional input to control its state, thus mimicking the control of biological synapses by neuromodulators.
This study validates the use of stress as a new way of controlling analog memristors and this stress-based approach opens up a multitude of new opportunities for the operation of synaptic networks.

1 K. Rim, J. L. Hoyt, and J. F. Gibbons, International Electron Devices Meeting, 1998
2 J. L. Hoyt et al., International Electron Devices Meeting, 2002
3 H. Jeong, and L. Shi, J. Phys. D: Appl. Phys., 2019
4 J. Fompeyrine, Y. Popoff, S. Abel, US20210034320A1, 2021
5 L. Bégon-Lours, M. Halter, Y. Popoff, Z. Yu, D. F. Falcone, and B. J. Offrein, ESSDERC, 2021
6 J. Fompeyrine, Y. Popoff, S. Abel, US2020235294A1, 2020

Neurodegenerative diseases and brain damage require the constant development of new technologies able to replicate synaptic functionalities and replace damaged connections. As widely reported in literature, synaptic communication is based on electrochemical mechanisms: certain stimuli might induce the release of neurotransmitters from the pre-synaptic site and the consequent recognition of these molecules at the post-synaptic receptors, thus causing the propagation of an action potential from one cell to another.¹ In this scenario, neuromorphic devices represent a powerful platform to recreate artificial neural networks, which will offer the chance to improve implantable devices towards the replacement and repair of injured brain areas. In particular, organic neuromorphic devices, such as PEDOT:PSS organic electrochemical transistor (OECT), offer a unique approach: in addition to the well-known biocompatibility of PEDOT:PSS, owing to its ionic-to-electronic current transduction, the conductance of these devices can be modulated through external stimuli (i.e. ionic flow) and the altered conductance state is retained over long time, replicating to some extent the physiological synaptic plasticity.² Recently we have successfully established a direct coupling between PC12 neuron-like cells, able to secrete dopamine (DA)³, and an organic neuromorphic device, thereby recreating an in vitro model of the synaptic system which simulate both long and short-term plasticity through the oxidation of DA.⁴
Here, we aim to recapitulate not only the same functionalities of neurons (i.e. synaptic plasticity) using an organic neuromorphic device, but also to have a biomimetic platform replicating the physiological neuronal membrane by means of supported lipid bilayers (SLBs), simplified in-vitro models of cell membranes with potential bio-technological applications ⁵. We evaluated the role of SLB composition on the OECT plasticity, analyzing how the bilayer fluidity affects the ions flow towards the neuromorphic channel, thus modulating the de-doping level of the PEDOT:PSS. Furthermore, we explored SLBs with neuronal composition in order to resemble the functioning of biological synapses. We expect that such platform thanks to its learning capabilities and biomimetic properties, could really interact with biological neurons, monitoring their activities and providing feedback accordingly, thus creating an artificial biohybrid functional network. This biohybrid platform might be further exploited to engineer future implantable devices able to replace damaged connections either in pathological networks in neurodegenerative diseases, or in case of amputations to create adaptive prosthetics able to reduce the mismatch between these artificial parts and the human body.


References.
1. Burns, M. E. & Augustine, G. J. Synaptic structure and function: Dynamic organization yields architectural precision. Cell 83, 187–194 (1995).
2. Gkoupidenis, P., Schaefer, N., Garlan, B. & Malliaras, G. G. Neuromorphic Functions in PEDOT:PSS Organic Electrochemical Transistors. Adv. Mater. 27, 7176–7180 (2015).
3. Westerink, R. H. S. & Ewing, A. G. The PC12 cell as model for neurosecretion. Acta Physiol. 192, 273–285 (2008).
4. Keene, S. T., Lubrano, C., Kazemzadeh, S., Melianas, A., Tuchman, Y., Polino, G., Scognamiglio, P., Cinà, L., Salleo, A., van de Burgt, Y., Santoro, F. A biohybrid synapse with neurotransmitter-mediated plasticity. Nat. Mater. (2020).
5. Richter, R. P., Bèrat, R. & Brisson, A. R. Formation of Solid Supported Lipid Bilayers an Integrated View. Langmuir 22, 3497–3505 (2006).

Functionalized sensor arrays are commonly used in the detection and sensing of chemical compounds in electronic tongue/nose systems. The analysis of the multivariate complex impedance data from sensor arrays is based on the global selectivity principle, which allows the qualitative discrimination of substances and their corresponding concentrations. However, frequently used analytic methods such as principal component analysis (PCA) and interactive document map (IDMAP) are non-quantitative methods and do not allow the development of calibratable devices. We propose a novel approach for the analysis of multivariate impedance data from sensor arrays used as electronic tongue devices. The method allows the representation of the data in a quantitative 2-D space where devices produced using similar fabrication parameters yields similar values when exposed to the same compound, at the same concentration. To prove the concept of the method, an extensive set of experiments was carried out, comprising the impedance (capacitance) vs. frequency spectra from 8 different types of electrodes (functionalized with organic semiconducting films), and 5 types of analytes, at different concentrations. Triplicated devices were also tested to confirm the reproducibility of the electrical impedance response. The result was a large, multivariate dataset, which required innovative computational aid to be analyzed. Although unspecific electrodes for the detection of substances were used in the experiments, several combinations of electrode arrays were possible to discriminate the analytes, with good reproducibility of the response. The concentration dependence of the electrode array response also permitted the determination of the detection limit for each compound, as well as each detection calibration curve. The method can be easily extended to smart sensing applications, including biosensors, which demand reliable and quantitative analysis.

Neuromorphic computing systems mimicking biological neural network show significant advantages in the efficiency for processing complex data. As a major building block of the new-generation computing hardware, artificial synapses such as those based on novel memory devices are capable to store and process information on the same device, mitigating the von Neuman bottleneck of digital systems featuring physical separation of memory and processing units and thus large time and energy overheads in running AI tasks. To harvest the advantage of in-memory computing, synapses are required to update their weights in a linear and repeatable way, and the operation is supposed to be fast and low power-consumption for in situ training, which remains one of the biggest challenges to various in-memory computing based neuromorphic systems. Besides, the long-term non-volatile memory property of the artificial synapse for retaining the weight value is critical to the learning performance. Here, we demonstrate a neuromorphic device by employing non-volatile organic electrochemical transistors (OECTs) as artificial synapses, which can provide analog weight updates with decent linearity while costing low power supply. The non-volatility of the synapse endows the possibility to emulate long-term synaptic plasticity, augmented by linear and repeatable weight modulation, provides a promising scheme to physically implement mainstream machine learning techniques such as gradient descent. This would significantly benefit the neuromorphic hardware in terms of power consumption, speed, form factors and cost. To verify this, we performed ANN simulations based on the weight update data of the transistors for the learning and recognition of the MNIST handwrite digits dataset. The results indicate the potential of this non-volatile OECT device for further moving towards smart and efficient edge neuromorphic computing hardware.

Resistive switching cells as one of the next-generation nonvolatile memories to replace the Flash have been extensively studied because of their potential use for neuromorphic devices. However, the mechanism of resistive switching and the origin of conductive filaments in resistance change materials remain controversial. At present, many combinations of electrode metals and resistance change materials have been reported. In particular, conductive filaments in conductive-bridge random access memories are formed by electrode metal such as silver (Ag) extending from the counter electrode instead of oxygen vacancies, and in several cases quantization phenomenon appears upon a voltage pulse application to the cell [1]. In this study, we focus on the effect of electrode materials on the forming and resistive switching characteristics in double tantalum oxide (Ta₂O₅)-based resistive switching cells to clarify the origin of the conductive filament created by the forming.
The resistive switching cells were fabricated as described below. A platinum (Pt) bottom electrode was prepared by dc sputtering. A 10-nm-thick Ta₂O₅ resistance change layer and a subsequent 10-nm-thick TaO_x (x=1.6) oxygen-vacancy reservoir layer were deposited by reactive rf sputtering with different oxygen gas flow rates [2]. Next, top electrodes (TEs), such as Pt, Ag, and gold (Au), were deposited through a metal mask by resistive heating (Ag and Au) or electron beam (Pt) evaporation. We examined the differences in initial resistance, forming voltage, and resistive switching characteristics of three types of cells, Ag, Au, and Pt-TE cells. The bottom electrode was grounded, and voltages were applied to the top electrodes with a diameter of 100 μm using a semiconductor parameter analyzer Keithley 2450. Current or voltage sweeps were applied to the top electrodes to investigate electrode material dependence of initial resistance and forming characteristics, or resistive switching characteristics, respectively.
The results of initial resistance were 5 k∼300 kΩ for the Ag-TE cells, 10 k∼200 kΩ for the Au-TE cells, and 2 k∼3 kΩ for the Pt-TE cells. Besides, the variations of forming voltage at which a cell resistance reaches sub-kΩ were 6~8 V, 7~10 V, and 5 V for the Ag, Au, and Pt-TE cells, respectively, indicating that the Pt-TE cells were more stable with less variation. These results remain unchanged even after a few months except for Ag-TE cells, which show a clear reduction in the forming voltage and the initial resistance. Although the reason for the reduction is unclear at present, Ag ions may be gradually dissolved from the Ag-TE layer to the Ta₂O₅ layer over time, owing not to the oxidation of metal-TE leading to an increase in the cell resistance. Note that the three types of cells after the specific forming show analog resistive switching characteristics upon bipolar voltage sweeps irrespective of the reduction above. These results reveal that an initial state in every cell does not involve the analog resistive switching.
Moreover, Ag and Au-TE cells exhibit discrete conductance fluctuation before the forming, while Pt-TE cells show continuous conductance increase [3]. The appearance frequency of the cell conductance based on quantized conductance is apparently high in Ag-TE cells. From viewpoint of the large variation of initial conductance and forming voltage, Ag ions can pass through the TaO_x layer with random ion collisions and form conductive filaments composed of Ag atoms during the conductance fluctuation and the forming, instead of oxygen vacancy rows supplied from the reservoir layer. Furthermore, in contrast, Pt-TE cells show possibly gradual diffusion of oxygen vacancies from the reservoir layer at the forming.

[1] T. Tsuruoka et al., Nanotechnology, 23, 435705 (2012).
[2] T. Miyatani et al., MRS Advances, 4, 2601 (2019).
[3] Y. Nishi et al., Journal of Applied Physics, 124, 152134 (2018).

This talk will provide an overview of our recent developments in polymer materials with unique electrochemical properties for use in neural bio-interfacing. We design, synthesise and fabricate conjugated polymer-based scaffolds presenting complex electrical and topographical cues that support neural stem cell adhesion, growth and differentiation [1]. I will discuss the applications of our facile strategy for controlled functionalisation and post-polymerisation of the aromatic backbone of conjugated polymers for selective cell stimulation [2]. Our recent developments in the context of bioelectronic nanostructured surfaces for biosensing and interfacing will be discussed [3,4].

[1] K. Ritzau-Reid… M. M. Stevens. “An electroactive oligo-EDOT platform for neural tissue engineering.” Advanced Functional Materials. 2020. DOI: 10.1002/adfm.202003710.
[2] A. Creamer… M. M. Stevens, M. Heeney. “Quantitative post-polymerisation functionalisation of conjugated polymer backbones and its application in multi-functionalised semiconducting polymer nanoparticles.” Nature Communications. 2018. 9: 3237.
[3] S. Higgins… M. M. Stevens. "High-aspect-ratio nanostructured surfaces as biological metamaterials." Advanced Materials. 2019. DOI: 10.1002/adma.201903862.
[4] C. Chiappini… M. M. Stevens, E. Tasciotti. “Biodegradable silicon nanoneedles delivering nucleic acids intracellularly induce localized in vivo neovascularization.” Nature Materials. 2015. 14: 532-539.

The seamless integration of biology with electronics demands for new bio-inspired approaches able to detect, process and amplify both ionic and electronic information. Ions are the ubiquitous biological and physiological regulators, being involved in most of the fundamental processes of every living organism. Ions enable the communication between cells and the metabolic and bioenergetic processes, playing a key role in hydration, pH regulation, and osmotic balance across cell membranes. Biological systems, including animals and plants, communicate and process information using as carriers ions, small molecules, and electronic charges. In the nervous system, stimuli are collected from distributed sensory receptors, computation then takes place locally or centrally, and when necessary, feedback signals drive sensorimotor processes. The technological potential of co-integration between electronics and biology stems from the ability of electronics to bi-directionally interact with biological systems, and even more importantly to emulate biological functions.^[1,2]
Using state of art organic electrochemical materials and devices, we show the design and validation of unconventional iontronic circuit and system approaches for (i) integrated ion-sensing amplifier and (ii) spatiotemporal dynamic multiplexing of ionic signals in a shared liquid medium of communication. The ion-sensing amplifier integrates both selective ion-to-electron transduction and local signal amplification demonstrating a sensitivity larger than 2300 mV V⁻¹ dec⁻¹, which overcomes the fundamental limit.^[3] It provides both ion detection over a range of five orders of magnitude and real-time monitoring of variations two orders of magnitude lower than the detected concentration, viz. multiscale ion detection. The iontronic multiplexing system discriminates locally random-access events with no need of peripheral circuitry or address assignment, thus deceasing significantly the integration complexity.^[4] The form factors allow for intimate bio-interfacing communication in shared medium. The ion-sensing amplifier and the spatiotemporal dynamic iontronic multiplexer are fundamental building blocks for multifunctional bio-interfacing and sensing in the emerging fields of bioelectronics, wearables, and neuromorphic computing or sensing.

[1] S. Vassanelli, M. Mahmud, Front. Neurosci. 2016, 10, 438.
[2] C. Lubrano, G. M. Matrone, C. Forro, Z. Jahed, A. Offenhaeusser, A. Salleo, B. Cui, F. Santoro, MRS Commun. 2020, 10, 398.
[3] P. Romele, P. Gkoupidenis, D. A. Koutsouras, K. Lieberth, Zs. M. Kovacs-Vajna, P. W. M. Blom, F. Torricelli, Nat. Commun. 2020, 11, 3743.
[4] D. A. Koutsouras, M. Hassanpour Amiri, P. W. M. Blom, F. Torricelli, K. Asadi, P. Gkoupidenis Adv. Funct. Mater. 2021, 31, 2011013.

Increasing number of electrodes in neural recording and stimulation devices is a key aspect to improve the spatial resolution. To keep appropriate dimensions for clinical applications, utilization of multiple layers of superimposed conductive leads helps to maximize the number of electrodes in a limited available space. However, close conductive lines with different electrical charges can lead to a crosstalk effect between channels due to capacitive coupling. To understand and quantify this phenomenon on devices made of gold leads with Parylene C encapsulation, we present a study on the influence of the thickness of inter-conductive insulation layer on capacitive coupling and Parylene C permeability. Subsequently, in vitro experiments were conducted on a multielectrode array, consisting of spatially distributed multi conductive-layer electrode clusters. This helped to quantify capacitive coupling. Furthermore, in vivo experiments were carried out to measure crosstalk in multi conductive layer electrocorticography probes, interfacing the brain. Finally, by shedding light on the optimal insulation-layer thickness between conductive layers, this study will help to optimize organic-bioelectronic sensor’s structure, to increase number of electrodes, thus spatial resolution of devices. Therefore, this will lead to an efficient detection of biological events and facilitate interfacing of these devices with small anatomical structures.

The emerging field of organic electrochemical neuromorphic devices is gaining enormous interest as such devices show distinct synaptic properties, operate at a low voltage, and enable biocompatible, intelligent sensor units which might be used, e.g., for real-time classification of bio-signals. So far the focus of research is mainly on the synaptic properties of single organic electrochemical transistors and related devices with the emphasis to control the strength and time constants of synaptic plasticity. However, to do computation, e.g., in a feedforward neural network, a very large number of identical synaptic devices is required with very strict definitions in terms of their synaptic properties to enable efficient learning. Due to these challenges, organic electrochemical synaptic networks have not been used for complex classification tasks.
In this contribution, I discuss how networks of synaptic organic electrochemical devices can be employed to solve complex classification tasks by utilizing a brain-inspired approach of computation denoted as echo-state or reservoir computing. Reservoir computing, a kind of recurrent neural network, is an approach where a low-dimensional input signal is mapped onto a high-dimensional output through a non-linear transformation which exponentially increases the probability to classify the input data according to the features being selected. The reservoir itself does not require training which makes this technique very fast and power-efficient compared to other machine learning approaches.
We create such a reservoir by growing dendritic networks of the benchmark material PEDOT:PF6 in an array of electrodes using field-directed electropolymerization. This method enables us to grow random networks of conductive polymer fibers with a diameter down to 1µm. As these fibers, each operating as an organic electrochemical transistor, are connected through the liquid electrolyte, a situation of complex resistive and capacitive coupling is created in the network. Hence, if an input signal is applied to the reservoir, the signal is distorted in a highly non-linear and transmitted to the output. The degree of non-linearity of the polymer reservoir depends on the balance between inhibitory and excitatory fibers and the strength of capacitive coupling. Implementing a delayed feedback loop connecting input and output of the reservoir (so-called single-node reservoir), the complexity of the system at hand is proven by chaotic oscillations. Ultimately, the single-node reservoir is utilized to demonstrate standardized classification tasks such as heartbeat classification, classification of flowers, time-series prediction, and others.

Brain-inspired neural networks are emerging as potential alternatives to the traditional von-Neumann architectures, mainly due to their unique structure of combining the memory and processor, giving them the ability to learn more efficiently ^[1]. Such neural architectures are composed of computing and memory elements, representing the neuron and synapse functions, respectively. Neurons are used to execute current summation, integration, and firing. As frequent integrating and firing events are involved, the power consumption needs to be low enough to ensure an energy-efficient computing element. Currently, neuronal functionality is often performed using a conventional CMOS-based neuron. However, a CMOS neuron, made of several transistors and a capacitor, is area- and energy-inefficient ^[2]. Hence, there has been a great interest in searching for an alternative neuron device by deploying new concepts, designs, and materials. Recently, a compact neuron design has been exploited utilizing two-terminal threshold switching (TS)-based elements, owing unique features such as thresholding and self-relaxation behaviors.
In our previous work, we developed Ag/HfO₂/Pt-based volatile electrochemical metallization (v-ECM) devices ^[3]. The stack was built from atomic layer deposited (ALD) 3 nm HfO₂ sandwiched between an Ag top electrode and a Pt bottom electrode. The thin HfO₂ layer serves as electrolyte matrix for the formation/rupture of an Ag filament. The fabricated device exhibits an extremely low-leakage current (I_off) of < 100 fA and ultra-low threshold voltage (V_th) of < 0.2 V, guaranteeing a low power consumption. In most threshold switching devices such as ovonic threshold switches (OTS) and insulator-to-metal transition (IMT) devices, the I_off and V_th show a tradeoff, meaning that the improvement in leakage current comes with increased V_th and vice versa. In this regard, v-ECMs have a unique advantage of providing low I_off and low V_th values simultaneously.
Some of the main concerns in such devices are the volatility range (stuck-at-ON) and the switching variability. The morphology of the conductive filament ultimately defines how fast the device relaxes back to the initial state. Excessive Ag in the filament could lead to a non-volatile state. Often, this happens during the initial forming stage, since a relatively higher voltage is required to initiate the switching. Therefore, controlling the amount of Ag constituting the filament is important. The formation/rupture of conductive filaments in v-ECM is closely related to the electrochemical dynamics of the active metal, such as Ag, which could determine the switching speed and the morphology of the filament. In addition to the material’s property, the electrochemical dynamics can be modulated by external factors like the electric field, bias duration, and temperature.
In this presentation, we will discuss the advantages of Ag/HfO₂/Pt based TS for artificial neuron applications, and potential improvements to enhance performance such as variability, volatility, and switching speed. For this purpose, different approaches to control the Ag concentration in the matrix - such as hot forming and annealing - will be discussed. Finally, a proof of concept artificial spiking neuron functionality of our device will be shown.

B. Chen, et. al., “Efficient in-memory computing architecture based on crossbar arrays,” 2015 Ieee International Electron Devices Meeting (IEDM), 2015, pp.
G. Indiveri, et. al., “Neuromorphic Silicon Neuron Circuits,” Frontiers in Neuroscience, vol. 5, pp. 73, 2011.
S. A. Chekol, et. al., "An Ag/HfO₂/Pt Threshold Switching Device with an Ultra-Low Leakage (< 10 fA), High On/Off Ratio (> 10¹¹), and Low Threshold Voltage (< 0.2 V) for Energy-Efficient Neuromorphic Computing," 2021 IEEE International Memory Workshop (IMW), 2021, pp. 1-4.

Non-associative learning represents fundamental mechanisms by which organisms interact with their environment. Early studies in neuroscience using sea slugs as model systems demonstrated the synaptic origins of habituation and sensitization. In this presentation, we will discuss non-associative learning mechanisms in NiO, a canonical Mott insulator. By varying the Ni:O stoichiometry, it is possible to obtain a range of transport gaps as well as p-type conduction. Using different stimuli such as oxygen chemical potential and UV light, we have been able to obtain current-voltage characteristics (resistance evolution) that emulate cellular learning in two-terminal NiO devices. Spectroscopic evidence on representative samples indicates that modulation of oxygen vacancy concentration through the film thickness is responsible for such non-associative learning. A set of samples carefully prepared under different oxygen partial pressures (by varying Ar-O₂ ratio) have been studied for resistive switching and learning under electric field pulsing. We find that samples with near-stoichiometric composition do not show switching behavior. However, interestingly, we are able to control the memory formation dynamics in oxygen-deficient samples by training under similar electric field strengths. We have studied such samples in-depth using X-ray scattering techniques and performed drift-diffusion modeling of current-voltage curves to understand the mechanisms behind learning and glass-like relaxation; and will be discussed in this presentation. Mott materials with carefully chosen defect structures possess highly tunable electronic structures and appear to be promising candidates for neuromorphic computation.