Novel Approaches for Neuromorphic Computing: Materials, Concepts and Devices

Resume : The polarization reversal in ferroelectric HfO2 films is adopted to store information in three distinct device classes. Depending on the stack composition different devices can be constructed from the very same ferroelectric layer - ferroelectric capacitors (FeCAP), ferroelectric field effect transistors (FeFET) and ferroelectric tunnel junctions (FTJ). That is, the electrical characteristics of this devices are strongly influenced by the whole material stack, rather than being dictated by the properties of the ferroelectric layer itself. In my talk I will discuss the different device concepts and their suitability for memory and for beyond-memory application such as logic-in-memory or neuromorphic computation.

Resume : Phase Change Materials (PCM) and ferroelectric RAMs have been investigated to achieve chips with neuromorphic hardware capable of processing larger amounts of information. PCM are usually based on GeTe and GeSbTe. GeTe has recently gathered renewed interest for its ferroelectricity, whose control would enable a new regime of usage, including the non-volatile control of the spin transport [1]. Here we exploit GeTe thin films in their ferroelectric (FE) regime, by electrical gating. The switching of the polarization through voltage pulses is measured as resistance variation of metal/GeTe heterojunctions (M/GeTe). The modulation of resistance is due to the different local band bending induced by the screening of the polarization charge. Piezoresponse Force Microscopy correlated the microscopic distribution of FE domains with the electrical resistivity of the junction. The low-voltage control is provided by the coercive voltage (3-7 V), with modulation of resistivity up to 800%. The switching is robust, with endurance up to 105 cycles. The number of subsequent voltage pulses, their amplitude, and duration, as well as exploiting ferroelectric minor loops, enable a continuous distribution of intermediate resistive states. The memristive behavior of GeTe in its ferroelectric regime holds potential for the development of a technology platform for neuromorphic computing based on materials “on the shelf” of semiconductor industry. [1] C. Rinaldi et al., Nano Letters 18, 2751 (2018)

Resume : Abstract: HfO2 based ferroelectric devices have attracted increasing attentions as a promising candidate for future information processing technology applications due to the excellent characteristics such as great scaling ability, good CMOS compatibility and high remnant polarization (Pr) [1-3]. It is critical to deeply understand the cycling behaviors and the correlated physical effects. In this work, the HfO2-based ferroelectric devices with the structure of TiN/Hf0.5Zr0.5O2/TiN were fabricated under different post annealing temperatures. The cycling behaviors of HfO2-based ferroelectric devices are measured under various operation schemes with different pulse waveforms. The measured datas show that the cycling behaviors during wake-up and fatigue processes are correlated with the pulse waveform. Meanwhile, the dependence of the cycling behaviors on the post annealing temperatures are also observed. These observed cycling behaviors are correlated with the generation and redistribution of oxygen vacancies in the Hf0.5Zr0.5O2 layer by using the physical mechanism [4]. References： [1] T. S. Böscke et al., “Ferroelectricity in hafnium oxide thin films”, Appl. Phys. Lett., vol. 99, no. 10, pp. 102903, 2011. [2] J. Müller et al., “Ferroelectricity in HfO2 enables nonvolatile data storage in 28 nm HKMG”, VLSIT 2012”, VLSIT 2012, pp. 25–26. [3] J. Müller et al., “Ferroelectric hafnium oxide: A CMOS-compatible and highly scalable approach to future ferroelectric memories”, IEDM 2013, 10.8. [4] C. Liu et al., “Role of Oxygen Vacancies in Electric Field Cycling Behaviors of Ferroelectric Hafnium Oxide”, IEDM 2018, 16.4.

Resume : The emergence of ferroelectricity in doped HfO2 has attracted a great deal of attention since its discovery in 2011. Among the various compositions, HfxZr1-xO2 (HZO) films have emerged as one of the more promising material. The ferroelectricity in thin doped HfO2 films is generally accepted to originate from the crystallization of a non centro-symmetrical orthorhombic phase. Here we report the utilization of millisecond Flash Lamp Annealing (ms-FLA) for the stabilization of this ferroelectric phase, down to a thickness of 3 nm. The combination of a 120 s long preheat step of 375°C with a 20 ms flash lamp pulse results in ferroelectric HZO films with TiN electrodes. This statement is supported by x-ray diffraction (XRD), polarization hysteresis and piezo force microscopy (PFM) measurements. In 10 nm and 5 nm thick HZO layers a remanent polarization (Pr) of ~21 µC/cm2 and ~15 µC/cm2 is achieved, respectively. Cycling analysis shows an increase of endurance for the ms-FLA compared to rapid thermal annealing (RTA) by one decade, up to 107 bipolar cycles. An SiO2 encapsulation further increases the endurance to 108 cycles. For the 3nm thin HZO sample, the preheat temperature of the ms-FLA was increased to 550°C to successfully crystallize HZO in the orthorhombic phase. On the contrary with standard RTA anneals, no conditions could be found for 3nm thick films without a fraction of the layer crystallized in the monoclinic phase. To our knowledge, this is also the first time XRD data show clear evidence for the orthorhombic phase in 3nm films. We also demonstrate the advantage of this approach using other electrodes than TiN. This shows the benefits of a fast thermal process to stabilize ultra-thin ferroelectric films, which are promising candidates for non-volatile memory and neuromorphic hardware applications when used as tunnel junctions.

Resume : Organic memristive device was designed for mimicking some properties of synapses. We will consider several applications of these devices for the realization of perceptrons, hardware realization of models of parts of animal nervous systems, frequency dependent plasticity, and STDP learning. Finally, a results of the use of these devices for connecting two neuronal cells from the rat cortex will be discussed.

Resume : Artificial intelligence (AI) has been heralded as the flagbearer of the fourth industrial revolution. But it comes with a cost and that is computing power. It is projected that by 2040, we will need more computing energy than the total energy we can produce at the moment. So, we need devices that can offer higher computing/ storage density with low energy consumption. We are addressing these challenges using a molecular-electronic route. Historically, organic electronic devices have stimulated scientific excitements in optoelectronics, especially OLEDs but are yet to make any other significant technological impact. The reasons behind their limited success are their poor robustness, stability, endurance and most importantly, the lack of mechanistic understanding that restricts the emergence of approaches to solve these problems. We have overcome each of these difficulties in our memristors based on transition metal complexes of azo-aromatic ligands that exhibit high reproducibility (~350 devices), fast switching (?30 ns), excellent endurance (~10^12 cycles), stability (>10^6 s) and scalability (down to ~60nm2)[1, 2]. Using in-situ Raman spectroscopy we are able to track the electronic changes in molecules in-operando at every point of our voltage sweep providing a clear picture of our molecular mechanism that enables us to do different molecular and device engineering to achieve targeted functionalities. Using devices of this genre we are addressing the existing computing challenges via three routes, (i) By designing devices with ultra-low power: We are able to design memristors with switching voltage as low as 70mV, with energy ~36aJ/ 60nm^2. The current and voltage levels of these devices meet the requirements specified in ITRS road map. (ii) By designing memristors and memcapacitors with multiple discrete plateaus: We have developed memristors with 3- 4 distinct conducting plateaus which also shows mem-capacitance. Their concomitant occurrence is enabled by symmetry breaking of our film-molecules driven by voltage, a new paradigm in condensed matter physics. (iii) Brain-inspired computing: Using an unprecedented self-assembly of our molecular films we are able to realize discrete conductance plateaus in a memristor with sharp switching. We are able to demonstrate up to 29 well defined, reproducible, discrete states each of which are characterized by in-situ Raman. We are exploring the potential of this system to enable synaptic functionalities. References 1. Goswami S, Matula AJ, Rath SP, Hedström S, Saha S, Annamalai M, et al. Robust resistive memory devices using solution-processable metal-coordinated azo aromatics. Nature materials 2017, 16(12): 1216. 2. Valov I, Kozicki M. Non-volatile memories: Organic memristors come of age. Nature materials 2017, 16(12): 1170.

Resume : Recently, bio-inspired sensory devices are attracted a great attention, however, still they are limited in challenges in emulating the preprocessing abilities of sensory organs such as reception, filtering, adaptation, and sensory memory at the device level itself. Here, inspired by structure and parallel functions of reception and preprocessing the information of Merkel-cells, which form synapse-like connections with afferent neuron(Merkel-cell Neurite Complex, MCNC), we designed flexible and intrinsically intelligent tactile sensor. Using an organic synaptic transistor structure with ferroelectric nanocomposite gate dielectric of barium titanate nanoparticles and poly(vinylidene fluoride-trifluoroethylene), synaptic functions are demonstrated. Furthermore, tactile stimuli induces triboelectric-capacitive coupling effect and ferroelectric dipoles switching so that post synaptic current of this architecture is expressed as tactile information with synaptic functions. Adaptation, filtering and memory functions can be achieved in a single device without any extra components, with synaptic functions which also can be modulated by varying the nanocomposite composition. This work provides a new paradigm of intelligent sensor for artificial perception systems by imitation of synaptic functions extended on function of slowly adapting reception, with simple structure of Merkel-cell neurite complex.

Resume : Poly(vinylidene fluoride) (PVDF) homopolymer and its copolymers with trifluoroethylene (P(VDF-TrFE)) are robust organic ferroelectrics which can be easily deposited onto any solid substrates with a good Van der Waals interfacial bonding on a large scale. We demonstrate here that even single-layer thickness organic homopolymer thin films remain ferroelectric at room temperature and can be used for ferroelectric tunnel junctions (FTJs). A TJ array based on PVDF homopolymer is designed on a Si substrate. We evidence the intimate link between polarization switching and resistance switching in these 190 nm wide solid-state FTJs, with a resulting room temperature tunnel electroresistance of more than 1,000%.[1] Further, by covering the PVDF copolymers on 2D MoS2, the ferroelectric transistor is demonstrated as a regulable artificial synapse in which the purely electronic mechanism associating with ferroelectric domain dynamics results in a steerable resistive switching within more than 1000 intermediate resistance states and as large as 4 orders. Essential synaptic plasticity, such as the long-term potentiation and depression (LTP/D) and spike-timing dependent plasticity (STDP), were emulated in a single device. The energy consumption for each synaptic operation is estimated to be less than 1 fJ. The endurance behaviour implies a 10-years life of a device working continuously at the biological brain work-frequency of 10 Hz. These performance gains and efficiency in the CMOS-compatible organic ferroelectric synapses [2] open a path towards implementation of unsupervised learning in high-density memristive crossbar arrays. References [1] B. Tian et al., Nat. Commun. 7, 11502, (2016) [2] B. Tian et al., Adv. Electron. Mater. 1800600 (2019)

Resume : The rise of big data creates a demand for novel computing architectures capable of solving hard problems with a large number of variables directly within the memory. We have recently introduced a novel concept of solving matrix algebra problems within memory arrays, where computing (multiplication and summation) is done physically within a crosspoint array of analogue resistive switching memory (RRAM) devices with close-loop configuration. Here, I will present the concept and illustrate the various implication for solving real-life problems, such as page ranking and differential equation. Scaling opportunities and a comparison with alternate computing concept, including digital, photonic and quantum technologies, will be discussed.

Resume : Emerging materials, novel technologies and unconventional computing approaches are promising replacements of conventional general purpose processors for the demanding neuromorphic applications, as they provide application-specific solutions to the physical limitations introduced by the CMOS technology and von Neumann computers. In this work, the memristor technology is combined with the segmented crossbar architecture for the exploration of novel hardware accelerators for neuromorphic applications. This combination exploits the high computational performance and low power dissipation of memristor networks with the integration scaling of the segmented crossbar, where the sneak path problem is suppressed. As a case study, text recognition based on Artificial Neural Networks with the proposed approach is presented and is also validated through SPICE-level simulations.

Resume : Memristor technology will strongly influence the architecture of computer systems in the near future. Its potential in several application domains, e.g. in-memory information processing [1], neuromorphic computing [2], hardware cryptography [3], and machine learning makes it more than ever necessary to understand the underlying resistive switching mechanisms and to look for electroforming-free memristors. We have developed an electroforming-free bipolar and an electroforming-free unipolar memristor, namely BiFeO3 and YMnO3, respectively. Impedance data analysis and quasi-static test measurements on BiFeO3 [4] with mobile oxygen vacancies and substitutional Ti donors on Fe lattice sites reveal that the field-accelerated ion mobility constitutes a source of ultra-nonlinearity and that the redistribution of oxygen vacancies during the writing step causes a nonvolatile reconfiguration of the barrier height of the top and bottom electrode. Temperature dependent transport measurements on YMnO3 [5] with vortex states reveal that the density of vortex states is changed during the writing step. A large and small concentration of vortices sets the YMnO3 into low resistance state and high resistance state, respectively. We discuss how Fuzzy logics are realized with electroforming-free memristors and how this can be exploited for machine learning. [1] You et al., Adv. Funct. Mat. 24, 3357-3365, 2014. [2] Du et al., Front. Neurosci. 9, 227, 2015. [3] Du et al., J. Appl. Phys. 115, 124501, 2014. [4] Du et al., Phys. Rev. Applied 10, 054025, 2018. [5] Rayapati et al. Journal of Applied Physics 124, 144102, 2018.

Resume : Resistive random-access memory (RRAM) is one of the most promising candidates for the realisation of synaptic weights in physically implemented artificial neural networks (ANNs). The main concern is that this might not be feasible due to the non-idealities of RRAM devices. In our work, we simulate the inference accuracy of physical ANNs that are implemented using non-ideal devices and consider each non-ideality separately in order to evaluate their effect on ANN performance. We analyse the effects of ratio of the high and low resistance states (HRS/LRS) in proportional mapping scheme, devices not able to electroform or stuck in one of the resistance states, current/voltage non-linearities, non-linearity of resistance modulation using voltage pulses and device-to-device variability. We find that in realistic scenarios HRS/LRS ratio and device-to-device variations have the most significant impact on inference accuracy. By employing the concept of a committee machine, we show that all non-idealities can be dealt with by combining multiple non-ideal physical ANNs to produce a single output that results in significantly better inference accuracy. Although this method results in larger area and higher power consumption due to multiple ANNs used, its parallel nature improves the robustness of the system and does not increase the computation time.

Resume : In-Memory Computing exploiting memristive devices is an exciting road for implementing highly energy efficient neural networks. This vision is however challenged by the variability inherent to memristive devices, as the efficient implementation of in-memory computing does not allow error correction. In this work, we fabricated and tested a differential HfO2-based memory structure and its associated sense circuitry, which are ideal for in-memory computing. We show that our differential approach achieves the same reliability benefits as error correction, but without any CMOS overhead. We also show that it can naturally implement Binarized Deep Neural Networks, a class of deep neural networks discovered in 2016, which can achieve state-of-the-art performance with a highly reduced memory and logic footprint with regards to conventional artificial intelligence approaches. Our system is fully satisfactory for image and EEG signal classification applications, and can achieve extremely high energy efficiency. Finally, we evidence how the extra reliability provided by the differential memory allows programming the devices in low voltage conditions, where they feature high endurance of billions of cycles, and showing a road for better exploiting the properties of Hafnium oxide based devices.

Resume : Magnetic materials have played a major role in information technologies. However, the word ‘magnetic’ refers in practice to one only class of magnetic materials known as ferromagnets which covered most of the data storage needs and are gradually entering random access memories, information processing and, more recently, neuromorphic computing. Antiferromagnets were discovered nearly contemporaneously with the transistor but despite being awarded a Nobel Prize in its Nobel Prize lecture, Louis Neel, admitted he found an interesting “but useless” magnetic order. Thus, antiferromagnets fell to academic research and only its auxiliary role in spin-valves brought some rather limited spotlight on them. In the past decade, antiferromagnets have been reborn while trying to improve spintronics effects on ferromagnets. It was found that most of the effects observed with ferromagnets had counterparts in antiferromagnets eventually with larger effects. Antiferromagnets started replicating many of the spintronics concepts based on ferromagnets and, by now, have established as an independent, mature, research field with some features that have been found to be exclusive for antiferromagnets. My contribution will focus on the recent most developments on the usage of antiferromagnetic materials on the field of neuromorphics. We will present the design of a completely analog spiking neuron build upon an antiferromagnetic thin film surrounded by standard CMOS transistors.

Resume : Deep neural network architectures have shown their superior performance in visual classification, audio recognition, signal recovery, astronomy information processing, etc. Their intrinsic parallel computation scheme can boost computation speed and reduce power consumption. Electronic neuromorphic computing has already been developed to reduce processor power consumption. However, the computing speed is limited by the interconnection electronic bandwidth. We propose to use photonic computation as an alternative to pushing computing speed further while exploiting brain-inspired architectures. We employ an Indium Phosphide (InP) based photonic neural network to demonstrate an all-optical two-layer feed-forward neural network. A non-linear semiconductor optical amplifier (SOA) is used to generate the optical nonlinear activation function: it is possible to change the shape of the nonlinear function by tuning the driven current and the power ratio of the CW control laser to the optical input signal. The combination of the on-chip photonic cross-connect for weighted addition implementation and the off-chip optical nonlinear function allows to process 10Gbit/s data sequences with a normalized root mean square error of 0.21 at the very final output, which is comparable to using an E/O conversion approach after each layer. This result paves the way to implement a reconfigurable all-optical deep neural network with photonic integrated circuits with co-integrated optical amplifiers.

Resume : Self-assembled conductive and memristive networks could play a role in future cognitive computation, for example in a winner-takes-all system or for reservoir computing. In epitaxial thin films of orthorhombic, antiferromagnetic TbMnO3 grown on SrTiO3, domain wall networks exist with domain walls that are perfectly vertical and atomically thin. In these walls, every other terbium atom is replaced by a manganese atom, leading to a chemical environment that is quite different from that of the domains and displaying ferromagnetic interactions localized at the walls. As it happens in other oxide thin films with ferroelastic domain walls, it is expected that the electrical properties of the walls also deviate from those of the bulk. Most often the conductivity enhancement is due to defect migration and local reduction of the Schottky barrier close to the domain walls. In the case of TbMnO3 on SrTiO3, the presence of ferromagnetism (below 40K) at the walls suggests an increase of their metallicity. Here we investigate the electrical properties of these films after growing them on conducting Nb-doped SrTiO3 substrates. The interface between the TbMnO3 film and the conductive substrate forms a p-n junction. We fabricate macroscopic 2-terminal and 4-terminal geometry electrodes to measure DC I-V curves and we fit the data to the Shockley diode equation including a resistance in series. We use a model for the resistance of the films, based on domain wall density, film thickness and other structural parameters, to obtain a value of the sheet resistance of these walls.

Resume : Acoustic sensing in the biological cochlea incorporates information processing at the sensor level, i.e. the hair cells, using the active dynamiccharacteristics of outer hair cells. These enable adaption of sensing properties to the actual hearing environment to improve frequency resolution, dynamic range and signal-to-noise-ratio and enable a selective amplification/suppression of signals, selected by attention focus or related to dangerous situations. We develop artificial hair cells based on micro-electromechanical systems, which provide sensing and neuromorphic pre-processing functionalities for acoustic signals. Therefore, our sensors are driven by a feedback algorithm into a nonlinear dynamics regime (near a bifurcation). Additionally, output-signal coupling is introduced to (i) increase the frequency range and (ii) enable nonlinear dynamics of the ensemble of coupled oscillators. Both regimes, the individual nonlinear oscillator as well as the coupled nonlinear oscillators, are used to realize a biomimetic processing. In particular, coupling increases the sensors response and shifts the system’s frequency range not only for the active but also for the passive case, thus providing an energy-efficient method for sensor adaptation. Due to the integrated filtering capability, the adaptable amplification and the operation principle based on coupled oscillators, this system can improve sensing technologies by reducing processing requirements and improving energy efficiency.

Resume : We have recently succeeded in the implementation of a large scale recurrent photonic neural network hosting up to 2025 photonic neurons. All net-work internal and readout connections are physically implemented with fully par-allel technology. Based on a digital micro-mirror array, we can train the Boolean readout weights using a greedy version of reinforcement learning. We find that the learning excellently converges. Furthermore, it appears to possess a conven-iently convex-like cost-function and demonstrates exceptional scalability of the learning effort with system size.

Resume : Hardware based neuromorphic computing, which requires a synaptic memory capable of retaining a multitude of addressable conductance states, opens a possibility to bypass the von Neumann bottleneck [1]. Ferroelectric field effect transistors (FeFETs) based on doped hafnium oxide have been demonstrated as viable candidates for neuromorphic synapses [2]. Here, the multitude of remnant polarization levels can be used to modulate the drain current. By utilizing any of three distinct types of signal sequences, the different levels can be addressed. Those sequences are described by a varying number of pulses, pulse width, or pulse height, respectively [3]. The resulting conductance response (CR) of the synapse can be described by a set of parameters: Abruptness, conductance variation and conductance dynamic range [4]. Here, we present the CR modulation capability of HfO2 based FeFETs. On the one hand, influences from the signal sequence and read gate voltage were investigated. In case of the former, the pulse width and amplitude were varied, whereas latter is demonstrated to vary the abruptness of depression and potentiation in opposed directions. On the other hand, influences resulting from the dopant, interface layer, and other device integration related quantities were investigated. Furthermore, endurance was investigated in regard to neuromorphic application. Additionally, it was demonstrated, that a CR with a nonlinearity coefficient smaller than 0.1 can be achieved for read voltages as low as 0.3V. [1] Indiveri, G. et al.; Proc. IEEE, 1379-1397, 2015 [2] Jerry, M. et al.; IEDM, 6.2.1-6.2.4, 2017 [3] Oh, S. et al.; IEEE Electron Device Lett. 38 (6), 732-735, 2017 [4] Gi, S.-G. et al.; IEEE Trans. Electron Devices 65 (9), 3996-4003, 2018

Resume : Motivated by the increasing interest in reconfigurable and neuromorphic computing architectures, we propose a reconfigurable logic gate able to reproduce several functionalities: OR, AND, XOR and CNOT [1]. This is achieved by manipulating the spin states in a double quantum wire with a coupling window enabled by a Rashba field. The ballistic spin polarized transmission functions are calculated using an effective mass scattering formalism applied to multi-channel, multi-terminal systems. For a given geometry and potential map of the embedded system, the device functionality can be switched by the external input only, set by the gate voltage, which tunes the Rashba coupling. Several configurations of the device region are investigated. Furthermore, different configurations of input and output terminals are considered. This approach shows that the proposed spintronic devices can be integrated in programmable architectures, implementing both classical and quantum algorithms. [1] G. A. Nemnes and Daniela Dragoman, Physica E 111, 13 (2019).

Resume : With the development of organic low power and biocompatible neuromorphic systems, novel paradigms based on integrating artificial neuronal network and machine learning with local adaptive biological systems can be envisioned. This approach could have applications in a variety of smart “trainable” devices, notably in adaptive prosthetics. To achieve a feedback mechanism between the biological systems and neuromorphic devices, the first critical step is understanding a complete spectrum of neuron to neuron communication. Current brain implants only focus on the recording of electrical signals from the action potentials of excited neurons. However, neuron to neuron communication is dominant by chemical recognition of neurotransmitter molecules. In this study, we succeed to bridge this gap for the first time by creating a neurotransmitter-sensitive organic neuromorphic device which directly translates chemical signals to a change in the neuromorphic memory state. Memory state of our neuromorphic device alters proportionally to different rates of dopamine secreted by PC12 neuron-like cells at the pre-synaptic. Applying voltage pulses at the pre-synaptic electrode drive the oxidation of dopamine at the electrolyte interface, leading to a change in synaptic weight at the post-synaptic electrode similar to short term and long term potentiation.

Resume : The term "computing" has a specific, firm, powerful, traditional meaning -- condensed in the paradigm of Turing computability (TC). A core aspect of TC is the perfectly reliable composition of perfectly identifiable symbolic tokens into complex, hierarchical symbolic structures. But all which is novel and promising and original in "neuromorphic" information processing leads away from such perfect symbolic compositionality. Apparently new formal conceptions of "computing" would be most welcome (and a new term for it, too). In my talk I will carve out a number of concrete aspects that separate neuromorphic information processing from symbolic computing - some of them classical topics in the philosophy of AI, others having more recently emerged from technological progress in non-digital hardware.

Resume : Neuromorphic analogue systems attract considerable attention to realize novel bio-inspired computing architectures. Circuits with advantageous features such as lower power dissipation and cognitive capabilities are main research goals. Double barrier memristive devices (DBMD) with the layer sequence Nb/Al/Al2O3/NbO𝑥/Au are promising candidates to emulate synaptic behavior in analog circuits. Selector-device free crossbar-arrays based on DBMDs have been already realized for pattern recognition tasks. The recognition performance of such systems strongly depends on the individual electrical I-V characteristics of the DBMDs and the deposition method of the layers, e.g. thin film magnetron sputtering. In this contribution, we show evidence that crucial parameters of the process plasma, such as floating potential, electron temperature as well as particle and energy flux at the substrate, are strongly correlated with I-V characteristics of the individual devices. These results are supported by transmission electron microscopy (TEM) and kinetic Monte Carlo simulations. Our findings enable a new pathway for the development of plasma engineered memristive devices.

Resume : Variability and noise are considered both an issue and an opportunity for hardware neural systems, depending on the target application. Indeed, deep learning hardware accelerators require high precision and reliability; in turn, variability and noise are fundamental, to some extent, for bio-inspired intelligent systems. Therefore, in view of the real deployment of the potential of RRAM devices for neural applications, their variability and noise characteristics must be investigated and understood. To this aim, we evidence that filamentary RRAMs produce telegraphic noise features of ionic origin when stimulated by trains of identical pulses.[1] The latter is the programming scheme of choice for RRAM neural applications. The observed zero-average telegraphic resistance variations are directly produced by pulses. The persistence of the resistance state after the delivery of a pulse is much longer than what is expected for conventional random telegraph noise (RTN). Indeed, RTN is dominated by charge trapping and release occurring within few tens of ms. In turn, the evidenced random resistance changes are stable up to 1 s, indicating that the phenomenon has an ionic origin. A model describing the stimulated noise amplitude as a function of the average resistance value is developed. The model considers a dynamic equilibrium condition at the interface between the conductive filament responsible for the switching and its interruption. [1] Brivio et al Scie. Rep. 9 6310 (2019)

Resume : Memristors as resistors with memory attract considerable attention due to their ability to mimic synapses in hardware neuromorphic systems (NS). Memristors based on nanocomposite (NC) (CoFeB)x(LiNbOy)100–x are good in this respect due to their high endurance, long retention time, and spike-timing-dependent plasticity (STDP). Resistive switching effect in such systems could be explained by conductive filaments formation and destruction caused by the electromigration of oxygen vacancies. NC film was synthesized by ion-beam sputtering in N2 ambient. The metal phase content x was ~6 at. %. Fabricated samples show excellent memristive characteristics: endurance >106 cycles, Roff/Ron > 100, retention time > 104 s, precision of resistive state programming better than 0.5%. We have observed a STDP behavior of NC memristors and experimentally demonstrate a simple way for the STDP window shape modulation by introducing the coefficients controlling the neuron spike amplitudes. In such a way the STDP window shape could be modulated from a classical asymmetric form to a bell-shaped, as well as to anti-STDP and to anti-bell-shaped. The possibility of STDP learning of NC memristor connected via analogue neurons based on op-amps was also demonstrated. Obtained results could be used for bio-inspired operation of future large memristor-based NS with associative and reinforcement learning ability. This work was supported by the Russian Science Foundation 16-19-10233.

Resume : In the vast flash-replacement emerging memory landscape Oxide-based resistive RAM (ReRAM) have been one of the most promising candidates in the last years and many works were published on the subject. Weebit-Nano is a company focused on SiOx-based ReRAMs. SiOx is a promising material thanks to its high band-gap, that allows to attain a clear separation among the resistance states, and its temperature stability that grants data In the vast flash-replacement emerging memory landscape Oxide-based resistive RAM (ReRAM) have been one of the most promising candidates in the last years and many works were published on the subject. Weebit-nano is a company focused on SiOx-based ReRAMs. SiOx is a promising material thanks to its high bandgap, that allows to attain a clear separation among the resistance states, and its temperature stability that grants data retention over time. In addition to its promising physical properties, SiOx is full CMOS compatible and offers high manufacturability. Finally, SiOx thickness and stoichiometry are easily tunable. By working on SiOx engineering and top electrode optimization we were able to tune the initial resistance of our devices (Rinit) and to correlate it with the maximum high resistance state (HRS) resistance. Measurements showed how there's a Rinit window around 100 [MΩ] that grants the highest possible HRS resistance and hence the best separation among memory states. A high enough memory window could also pave the way to the capability of encoding more than only 2 memory states between the low resistance state (LRS) and the higher HRS.

Resume : Self-assembled metal oxide nanowires (NWs) represent fascinating building blocks to artificially emulate human brain functionalities such as learning and memory, when their electronic properties are coupled with ionic effects [1,2]. Moreover, these low-dimensional systems offer the possibility to achieve new device functionalities by exploiting surface/quantum effects or by applying external stimuli such as illumination [2]. In this work, we propose a single crystalline ZnO NW as a model system to investigate ionic-related phenomena underlaying resistive switching and neuromorphic functionalities of ECM memristive cells. Indeed, by properly selecting materials and operating conditions, the single NW devices exhibited all memristive/neuromorphic functions such as non-volatile bipolar memory with multilevel capability, volatile switching and synaptic function through the emulation of Ca2 dynamics of biological synapses [3]. A detailed analysis of the physical mechanism of switching revealed the importance of interfaces in crystalline materials, since it was directly observed that ionic migration is restricted on the NW surface. Moreover, we show that ionic contribution from ambient moisture can modulate not only electronic properties of these devices but can also regulate the ionic transport mechanism. References [1] H.G. Manning et al. Nat. Commun. 9 (1), 3219, 2018 [2] G. Milano et al. Adv. Elect. Mater 1800909, 2019 [3] G. Milano et al. Nat. Commun. 9 (1), 5151, 2018

Resume : Memristors, non-linear devices of which the microscopic internal state can be modified and measured as a change in the conductance, can be used to emulate elements of the human brain, for example when used as synapses, they show synaptic plasticity through their ability to modulate the strength of connecting pathways. For this purpose Schottky junctions of ferromagnetic metals on Nb-doped SrTiO3, where memristive behaviour originates from electric field driven interfacial changes, are studied at room temperature. This behaviour is observed in as-fabricated devices, eliminating the need of an unfavourable electro-forming step. The resistive switching in these devices is gradual, allowing for well-controlled multilevel switching [1]. Some of the most important parameters influencing the resistive switching behaviour, including magnitude of applied voltage, time and number of applied pulses, are investigated. Different pulse widths and shapes gives rise to differences in the state-retention characteristics: in particular, longer lasting memory effects are realised by longer stimuli, which can be used to mimic learning and forgetting in short-term memory. Accumulative effects and how presenting multiple consecutive stimuli influences the overall state of the device are also investigated. The retention after the application of multiple pulses, in particular after a sequence of two pulses, shows that the last pulse in a sequence has the strongest influence on the observed retention characteristics. The direction of resistance change depends on the polarity of this final pulse, but earlier pulses are also shown to have a non-negligible influence. This could be used as a short term memory and to recall the last digit in a sequence. Application of a single negative (positive) pulse or sweep is seen to give rise to an increase (decrease) in resistance. When, on the other hand, repeat pulses of one polarity are sequentially applied, the observed changes in resistance are gradually enhanced. Using a power-law model, the positive pulsing action can be well-described. This modelling allows for the development of synaptic weight change learning rules, which can be used for training. Both for consecutive sweeping and pulsing measurements it is observed that the first event induces the largest resistance change and subsequent events give rise to increasingly smaller changes: this is reminiscent of the learning processes in the brain, where the largest change occurs during the first learning event. References [1] A. S. Goossens, et al., J. Appl. Phys., 124(15), 152102, 2018.

Resume : Operation of many of the industrial application is not possible without using superhard materials. Nowadays it is important to search for new cheap and effective materials which can substitute traditional materials in many field of science and technology. Traditionally material can be called as superhard if its Vickers hardness is higher than 40 GPa [1–3]. Here we predict new tungsten and molybdenum borides, some of which are promising hard materials that are expected to be thermodynamically stable in a wide range of conditions. We computed the composition-temperature phase diagram, which shows the stability ranges of all predicted phases. New boron-rich compound WB5 is predicted to be superhard with Vickers hardness of 45 GPa and at the same time it possesses high fracture toughness of ~4 MPa·m0.5 [4]. Newly predicted tungsten and molybdenum boride (WB5 and MoB5) are found to be thermodynamically stable in a wide range of temperatures at ambient pressure. Temperature dependences of the mechanical properties of WB5 were studied using quasiharmonic and anharmonic approximations. Our results suggest that WB5 and MoB5 remains a high-performance material even at very high temperatures. The work was supported by Russian Science Foundation (Grant No. 17-73-20038). Reference [1] V. L. Solozhenko, S. N. Dub, and N. V. Novikov, Diam. Relat. Mater. 10, 2228 (2001). [2] V. L. Solozhenko and E. Gregoryanz, Mater. Today 8, 44 (2005). [3] V. L. Solozhenko, O. O. Kurakevych, D. Andrault, Y. Le Godec, and M. Mezouar, Phys. Rev. Lett. 102, 015506 (2009). [4] A. G. Kvashnin, H. A. Zakaryan, C. Zhao, Y. Duan, Y. A. Kvashnina, C. Xie, H. Dong, and A. R. Oganov, J. Phys. Chem. Lett. 3470 (2018).

Resume : Recently, memristors have attracted considerable attention in the electronic devices community. Originally the memristor has been considered as the fourth fundamental two-terminal circuit element that provides a link between charge and magnetic flux, in addition to the resistor (link between current and voltage), the capacitor (link between voltage and charge), and the inductor (link between current and magnetic flux). Over the years, however, the definition of the memristor has changed, and today any two-terminal device showing a pinched hysteresis loop in its current-voltage characteristics is called memristor. Memristors are of particular interest since they possess an inherent memory effect and are able to emulate the function of biological synapses. In the present paper, we report on the fabrication and characterization of lateral few-layer MoS2 memristors as a specific type of memristive device. Our memristors consist of mechanically exfoliated MoS2 flakes transferred to a SiO2/Si substrate and two Ti/Au contacts serving as source and drain. The electrical characterization under different ambient conditions (air, N2 atmosphere, vacuum) reveals that both the shape of the hysteresis and the absolute current values strongly depend on the ambient. Moreover, it is shown that bare devices without protection (i.e., without a passivation layer) seriously degrade during operation, particularly in ambient air.

Resume : The prompt development of artificial intelligence systems and future advanced robotics incentives evolving demand on the understanding of the neuromorphic mechanism. Memristor is the fourth fundamental circuit element with a great opportunity for high-performance neuromorphic computing. We introduce a simple metal-insulator-metal system using a bioinspired material for memristor-based artificial synapse and demonstrate its behaviors. We emulated the physiological behavior of biological synapses by the evolution of the memristive device with external pulses. The plasticity of the artificial synapse is achieved, and it is comparable with the natural brain system. We represent a bioinspired neuromorphic device which is derived from a natural and waste material. The introduced memristor is capable to compete with its inorganic counterparts. We used an easy and inexpensive solution process to fabricate devices on a plastic substrate under ambient conditions. The electronic synapse is composed of collagen with magnesium electrodes that emulate synaptic functions by engineering the input pulses. In nature, neurons are communicated via synapses. Like a biological synapse (Figure 2), the top and bottom electrodes are regarded as the presynaptic and postsynaptic terminals, while the collagen thin film and cations are considered as the synaptic cleft and neurotransmitters, respectively. SEM image of the fabricated device confirms the well-deposited film and the desired structure of the neuristor. We have verified using a natural polymer to fabricate a biocompatible artificial synapse with flexibility, robustness, and analog conductance characteristics. The important synaptic characteristics including potentiation and depression, excitatory postsynaptic current (EPSC), paired-pulse facilitation (PPF), and spike-timing-dependent plasticity (STDP) and spike-rate dependent plasticity (SRDP) has been realized by the fabricated neuristor.

Resume : Neuromorphic computing, which emulates synaptic functions of biological neural networks, is emerging as one of the most viable successors of current computing paradigms. The use of established thin film deposition techniques, the safeguarding of Si compatibility, along with the precise control of synthesis, processing and characterization of multifunctional oxide films, is expected to mitigate current CMOS technology bottlenecks, through the production of two-terminal synaptic devices, structured in crossbar arrays. This work investigates the memristive/synaptic behavior and underlying mechanism of thin film cells based on LiCoO2-cathode in combination with various anode materials such as SiO2, SiO2/TiO2 and a:SiO2/poly-Si on p-doped Si substrates. For the fabrication of the multilayer thin film cells Pulsed Laser Deposition, conventional and reactive RF Sputtering have been employed. Depending on the anode material and thin film structure, reversible Li-ion migration from the LiCoO2 cathode, give rise to separate memristive functionalities such as binary or analog resistive switching, based on filamentary Li bridging and homogeneous insulator to metal transition (IMT) of LiCoO2 respectively. Additionally synaptic plasticity functions such as Paired Pulsed Facilitation (PPF) and Spike Timing Dependent Plasticity (STDP) along with the inherent stochastic feature of these devices disclose great potential for future spiking neural network applications.

Resume : Two dimensional materials such as graphene and hexagonal boron nitride have atomically thin structure. They can also stack and form a Van der Waals layered structure. For their atomic thickness, they have been studied a lot for the application to various electronic devices such as field effect transistors, photoelectric catalysts, and so on. Transition Metal Dichalcogenides (TMD) are one of the most popular two dimensional materials system for their various mechanical and electrical properties. Among them, MoS2 is the most familiar one for its abundance with stability and semiconducting band gap. However, TMD materials with Van der Waals layered structure have low electronic and ionic conductivity between the layers. Here we demonstrate resistive switching with vertically aligned MoS2 to minimize the loss of conductivity by the formation and rupture of conduction filament. Chemical Vapor Deposition (CVD) method is used for the synthesis of vertically aligned MoS2. Also Ab-initio Density Functional Theory (DFT) calculation is performed to verify the mechanism of conduction filament based on sulfur vacancies.