This talk describes state-of-the-art organic neuromorphic devices and provides an overview of the current challenges in the field and attempts to address them. I demonstrate two device concepts based on novel organic mixed-ionic electronic materials and show how we can use these devices in trainable biosensors, smart autonomous robotics and at the interface with biological systems. 

Next to that, the process of neural network training can be slow and energy-expensive due to the transfer of weight data between digital memory and processor chips. Neuromorphic systems can accelerate neural networks by performing multiply-accumulate operations in parallel using non-volatile analogue memory. However, the backpropagation training algorithm in multi-layer (deep) neural networks requires information - and thus storage - on the partial derivatives of the weight values, preventing easy implementation in hardware. 

In this talk I will highlight a novel hardware implementation of the well-established backpropagation algorithm that progressively updates each layer using in situ stochastic gradient descent, thus avoiding this storage requirement. We experimentally demonstrate the in situ error calculation and the proposed progressive backpropagation method using a multi-layer hardware implemented neural network based on organic EC-RAM, and confirm identical learning characteristics and classification performance compared to conventional backpropagation in software. 

Analog memristors without conductive filaments are promising candidates for neuromorphic and in-memory computing due to their superior multilevel switching capabilities, reproducibility, and low energy consumption. Among these, halide perovskite-based memristors—typically fabricated in metal–insulator–metal (MIM) architectures—leverage interfacial valence change mechanisms (VCM) or self-doping to modulate the Schottky barrier and achieve analog resistive switching. However, most reported devices exhibit highly non-linear current–voltage (I–V) characteristics, with a very narrow ohmic region, which limits their effectiveness in applications such as vector-matrix multiplication (VMM), where linear device response is essential for accurate and efficient computation.

Moreover, these devices often suffer from capacitive current spikes during transient measurements at low read voltages, which interfere with state detection. Although increasing the read voltage (>1 V) can suppress these transients, it risks disturbing the memristor’s programmed state—an undesirable tradeoff in practical applications.

In this work, we present a strategy to significantly broaden the linear/ohmic region in halide perovskite analog memristors through electrochemical doping induced by bias-driven metal ion migration from the top electrode. This results in an order-of-magnitude expansion of the ohmic window without compromising the analog switching characteristics. Structural and compositional analyses using scanning transmission electron microscopy with energy-dispersive X-ray spectroscopy (STEM-EDX) confirm the presence of migrated metal within the perovskite layer, while Kelvin probe force microscopy (KPFM) evidences the resulting local doping effect.

The engineered device exhibits stable and reproducible analog switching across 32 distinct conductance levels, each with retention exceeding 1000 seconds—key metrics for reliable in-memory computing. By combining filament-free operation, enhanced ohmicity, and robust multilevel retention, our approach addresses a fundamental limitation in halide perovskite memristors and paves the way for their integration into efficient and scalable neuromorphic architectures.

Metal halide perovskites are mixed ionic-electronic conductors. This mixed conduction allows them to be used for memristive applications. We have developed a cross-bar back-contacted device architecture for which we have demonstrated both artificial synapses and neurons with very low energy consumption.

However, perovskites are also excellent light absorbers. We show that when light is used to switch the state of the memristor, either in combination with voltage, or in addition to it, it can also be used to alter the properties of the memristor.

By using light as an input, perovskite memristors are ideally suited to perform in-sensor computation on visual input. We simulate such an application using the measurement parameters as input. We map the MNIST and N-MNIST datasets based on 4-bit inputs and train linear readout layers for classification. In this configuration, we find classification accuracies of up to 92.33 ± 0.06% and 84.34 ± 0.03%for MNIST and N-MNIST, respectively, with only minor deterioration by measurement noise. This result is more than 10% higher compared to a linear classifier for the N-MNIST dataset. The microscale device architecture lends itself well to high-density sensor arrays, ideally suited for efficient in-sensor computing.

In the pursuit of energy-efficient computing, various neuromorphic computing and engineering strategies have emerged. Silver nanowires (Ag NWs) are of particular interest for memristive applications, including neuromorphic architectures based on self-assembled nanowire networks [1].

Emergent behavior in these complex systems arises from resistive switching (RS) phenomena, occurring both within individual NWs and at their junctions. In the first case, this phenomenon is linked to rapture and rewiring of Ag NW by electromigration and/or Joule heating; the electrical connection can then be reestablished by forming of an conductive filament within the gap. In the second case, it is associated with the creation of conductive bridge at intersection of two NWs.

The current understanding of the phenomena underpinning the behavior of such networks is principally based on characterization of pre- and post-electrical stimulation devices and electrical measurements. However, such approaches lack direct insight into the local changes that might occur.

To address this, we performed in situ heating and biasing transmission electron microscopy (TEM) to directly visualize the structural and morphological evolution of single NWs and small NW networks under controlled electrical and thermal stimuli.

We investigated electrical breakdown in single Ag NWs during voltage sweep stimulation. Joule heating and electromigration are the two possible causes of fracture for metal wires during operation and in-situ observation sheds some more light on its mechanisms [2,3]. We also observed in situ the reformation of conductive pathways under applied bias, offering direct experimental evidence of self-healing behavior in metallic NW systems. This rewiring, involving the dissolution and redeposition of silver across the gap, highlights the potential for adaptive and reconfigurable network behavior.

Additionally, we studied the temperature influence on the morphology and structure of the single NWs and of small NW networks. In single Ag NWs, morphological changes were initially observed as the creation of humps and valleys, eventually with rising temperature the NWs break and the rapture continues to enlarge with changes in morphology following the crystalline orientations of preference.

In NW networks, progressive sintering at junctions was observed as the temperature increases. Further increase of the temperature led to formation of fractures, which progressively become larger, until the junction broke. In the context of neuromorphic network this suggests a failure and lose of its percolative pathways.

The hysteresis observed in perovskite solar cells sparked the interest in employing perovskites for memristive devices. Since the early reports, switching has been observed in all kinds of perovskite and perovskite-inspired materials sandwiched between metal electrodes. These devices have shown volatile as well as non-volatile behavior dependent on the system and the applied voltages. The switching has been explained by interfacial effects as well as filament formation, where various ionic species from the perovskite as well as the metal electrodes might be involved.

In this talk we will have a closer look at highest-performance perovskite memristors, which show on/off ratios larger than 10¹⁰ and excellent retention and endurance. Using various in-operando measurements such as infrared thermography, confocal photoluminescence, and electron microscopy, we elucidate formation and switching process in our memristors. Based on these insights, we present a novel optical patterning approach, which allows to down-scale the memristor area. Given the huge on-off ratio of these memristors, an outlook is given on other potential applications beyond neuromorphic computing.

The switchable bistable polarization in ferroelectrics allows for the binary control of optical, electronic, and catalytic properties that are essential for a wide range of applications. Going beyond the limitation of a binary remanent polarization and using that for evoking strongly non-linear material responses holds great promise for emerging spintronic and neuromorphic concepts. Here, we demonstrate that we can arbitrarily set the magnitude of the remanent ferroelectric polarization at the nanoscale in epitaxial PbZr_0.52Ti_0.48O₃ thin films with a single DC bias. By driving the ferroelectric system towards an instability near the PZT morphotropic phase boundary and controlling the resulting softness via epitaxial strain, we favor the formation of decoupled nanometric 180° domains that exhibit a broad coercive field distribution. Using in-situ optical second harmonic generation and X-ray diffraction, we investigate the emergence of the nanoscopic domain configuration. We then use piezoresponse force microscopy to demonstrate the possibility to locally and reversibly modulate the remanent polarization continuously between depolarized and saturated, while preserving the nanoscopic length scale of the domains. We highlight the direct technological relevance of nanoscale non-binary polarization switching, by showing first, the voltage-controlled tunability of the nonlinear optical response in our films and second, the quasi-continuous tunability of the tunnel electroresistance in ferroelectric tunnel junctions.

Introduction

The emergence of so-called artificial intelligence (AI) applications is creating lot of buzz not only in the scientific community, but also in public media coverage. AI applications have made huge leaps in recent years and are capable of human-like text creation, conversation, and forms of reasoning. However, they come at a huge energy cost, with some sources estimating tens of millions of kilowatt hours of electricity use per day, and some companies already consuming more electricity than many countries in the world.

One of the main reasons for these exploding energy costs is the reliance of AI hardware on conventional von Neumann architecture with a separation of memory and compute elements. Despite the immense improvements in GPU, algorithms, and software efficiency, some current large AI models rely on over a trillion model parameters, which are stored in off-chip memory and need to be moved back and forth constantly between memory and compute. This movement of data makes up the main energy consumption.

Neuromorphic in-memory computing holds immense promise to obviate this need for data movement. The most fundamental building blocks of this approach are crossbar arrays, where each intersection of a number of mutually perpendicular lines holds a combined memory and compute element, often referred to as memristor. Such memristor crossbar arrays could carry out vector-matrix multiplications, the basic operation of AI models, directly in hardware. There are several approaches to achieve such memristor functionality. All have their advantages and challenges, and for reasons of industry-compatible materials and simple device fabrication, leading to low cost, this work is focussed on resistive switching.

Hybrid resistive switching

Among resistive switching devices, the majority of implementations, especially in industry-compatible materials such as silicon oxide or hafnium oxide, is based on a reversible soft dielectric breakdown, called filamentary switching [1]. In this approach, it is very difficult to control multi-level resistance states due to the ultra-fast and ultra-nonlinear switching characteristics of the process. Consequently, this approach suffers from challenges of uniformity. As an alternative, resistance states can be controlled by the voltage-controlled redistribution of ionic species inside a switching film. This provides better uniformity and resistance control, but often suffers from poor state retention and the switching can be too slow.

Over the past few years, we have combined these two approaches into a materials design concept which we term “hybrid resistive switching”, where we combine the strengths of both approaches to overcome each other’s challenges. We did this by depositing self-assembled nano-engineered thin films at industry-friendly 400 °C, which switch by forming “partial filaments” inside the thin film, but leave an area close to the bottom electrode unperturbed, and thus form an effective switching interface. This provides finely controlled multi-level resistance states with excellent overall characteristics.

Here, we present the implementation of our hybrid resistive switching materials design concept with two example materials. One is based on sodium bismuth titanate [2], a mixed ionic/electronic conductor of great interest for a wide range of applications, and the other is based on industry-friendly hafnium oxide [3]. Overall, we demonstrate switching speeds down to 20 ns, over 500 separate and stable resistance states across several orders of magnitude, retention measured up to 300 days, and spike-timing-dependent plasticity as demonstration of neuromorphic functionality.

While at the materials and device level, the results are very promising to take forward and make into crossbar arrays as the initial and fundamental building block of future fast and energy-efficient bespoke AI hardware.

The potentiation and depression of synaptic conductivity regulate the plasticity and adaptability of synapses. In this discussion, we examine the general dynamic characteristics of ionic or electronic current conduction in memristors, which underpin the fundamental principles of synaptic activity. Key model requirements for memristors or chemical inductors to achieve conductance adaptation in response to incoming stimuli are outlined. We also propose various criteria, such as hysteresis and rectification, to achieve these properties. Additionally, we describe a range of diagnostic methods that link nonlinear time responses, the nonlinear cycling of current-voltage curves, and the linear frequency responses from impedance spectroscopy to evaluate adaptation properties. The frequency domain analysis of memristors and more generally, of conducting systems with memory features, provides essential information about the dynamic behaviour of the system.^1,2 The impedance response of a memristor can be represented as a linear circuit made of resistances, capacitors, and inductors, with voltage-dependent elements. We show the criteria that establish self-sustained oscillators for artificial neurons. The equivalent circuit properties allow to identify the Hopf bifurcation that produces a transition from quiescent to spiking regimes according to the incoming stimulus.

Materials that undergo a metal-to-insulator phase transition are interesting as threshold, or volatile, memristors, and they are a crucial component in self-oscillating circuits that emulate the emission of action potentials by neurons. Rare-earth (RE) nickelates (RENiO₃) show a metal-to-insulator transition at temperatures that can be tuned by different parameters, such as changing the RE cations, the strain state, film thickness or the oxygen vacancy content. Compared to other transition metal oxides, nickelates are especially interesting as memristive devices because of their endurance due to the robustness of the perovskite structure to high local temperatures and electric fields. In addition, it is known that the transport properties in transition metal (TM) oxides are largely dependent on the oxygen vacancy concentration. Thus, next to using the metal-insulator transition for artificial neurons, nickelates could also be used as synaptic devices driven by redox-reactions. However, this functionality is much less understood in the nickelates. Here we show that nickelates can behave as neuristors and as synapses and we discuss the mechanisms behind the different behaviours. In addition, we show that by interfacing nickelate thin films with ferroelectrics, it is possible to combine volatile and non-volatile memristive behaviour in one device and that such combination allows to select either the neuron or the synapse functionality by switching the ferroelectric polarization with the external bias[1]. Such devices could be used in reconfigurable networks, in which the devices can be dynamically reprogrammed to operate as memristors or spiking elements [2], for the implementation of sparse firing models in SNNs [3], or in oscillating neural networks that use memristive weights to couple the oscillations of individual elements [4]. These devices can also allow oscillators to be dynamically added or removed from a population.

.References

[1] R. Hamming Green. et al. Frontiers in Materials 11, 1356610, 2024
[2] R. John et al. Nat. Commun. 13, 2074, 2022
[3] G. Belec et al. Nat. Commun. 11, 3625, 2020
[4] P. Feketa et al. EEE Trans. Automatic Control 66, 3084, 2021

We report the development of ferroelectric and resistive memory arrays fabricated via atomic layer deposition (ALD) for neuromorphic and in-memory computing applications. A TiN/HfAlOx/Si-based ferroelectric memristor array exhibits intrinsic self-rectifying behavior, allowing selector-free crossbar integration. The device shows tunable short-term conductance decay characteristics, which are effectively exploited to implement physical reservoir computing (RC) for processing spatiotemporal signals. In parallel, a planar TiOx/Al₂O₃-based resistive memory array demonstrates analog switching with excellent linearity and endurance, enabling high-precision vector-matrix multiplication (VMM) operations with less than 2% error. To further enhance integration density, we developed a 3D vertical RRAM (VRRAM) array based on ALD deposited HfO₂  stacks, which provides reliable multi-level switching, high uniformity, and vertical scalability. The combination of 2D and 3D memory structures enables a compact, energy-efficient, and CMOS-compatible architecture for next-generation AI accelerators. These results underline the potential of hybrid memory arrays in realizing practical neuromorphic computing hardware with both learning and inference capabilities.

High-performance halide-based perovskite memory devices have been developed exhibiting a variety of synaptic [1-4] and neuronal functions based on non-volatile, and volatile or threshold switching, memristors, respectively. [5] However, a key ingredient in these perovskite-based systems is the presence of the highly toxic lead, which hinders their further development and commercial use. A lead-free perovskite approach for memristive applications could enable sustainable devices opening the path for practical applications, despite the current performance gap compared to lead-based systems. Herein, we present our recent data on the fabrication and characterization of printable non-volatile and volatile memristors based on Lead-Free Perovskites for artificial synapses and neurons emulation, respectively. Our approach is based on solution-processed manufacturing using all-inorganic, sustainable perovskites (Bismuth based) compounds. Depending on the metal contact type being either silver or gold, devices exhibit either non-volatile or volatile memristive switching. The non-volatile memristors exhibit an ON/OFF ratio of >10⁴ while demonstrating very good retention and cycling endurance characteristics exceeding 1000 seconds and 1000 cycles, respectively. Typical volatile devices exhibit an ON/OFF ratio of > 10³ and require a low switching voltages of few volts. Furthermore, linear long term potentiation protocols accompanied with an abrupt resistance suppression under depression protocols are demonstrated being also tunable by light illumination. The on-demand selection of the operation mode by tuning the metallic contact type, offers a unique materials system based on lead-free perovskites opening the path for implementing artificial synapses and neurons emulation in a single chip.

Halide perovskites possess mixed ionic and electronic conductivity, and the loosely bound halide ions leverage ion migration, which permits their application in memristors. Three-dimensional perovskites are plagued not only by their low moisture stability, but also by uncontrolled transport due to their polycrystalline nature that limits their application in memristive applications. Dimension reduction in perovskites allows improving the stability as compared to the bulk perovskites, and in this vein, layered perovskites are adaptable due to the wide choice of cation, and allow the tuning of microstructural and electrical properties. However, their large dielectric and quantum confinement limit their nonlinear conductance changes. To improve the neuromorphic device efficiency and training, it is paramount to achieve linear and symmetrical conductance changes through the Dion–Jacobson based layered perovskites. Vertically oriented layered perovskite-based synapses displayed a high device yield, low variation with synaptic weight storing capability, multi-level analogue states with long retention. Our developed vertically oriented perovskites eliminate the gaps between inorganic layers, which in turn allow the halide vacancies to migrate homogeneously regardless of grain boundaries to boost neuromorphic properties.

The traditional silicon-based computing systems are consuming excessive energy to process the huge amount of data generated daily. Neuromorphic computing is seen as a new, promising alternative as it can potentially reproduce the human brain’s efficiency.[1] Mixed Ionic-Electronic Conducting materials (MIECs) are ideal candidates to fabricate neuromorphic electronic devices as they can mimic the synapses in neurons. However, their performance and reliability need to be further improved to compete with the state-of-the-art silicon-based technology.

In this presentation, we will discuss diverse strategies that have been recently explored in our research group to improve the energy efficiency and performance of MIECs-based electronic devices, i.e., memristors and organic electrochemical memtransistors (OECmTs). Analyzing different examples, we will see how controlling the nature and environment of the migrating cations is a key strategy to tune synaptic plasticity and power consumption in these devices. In addition, we will introduce the use of a novel n-type organic MIEC to fabricate OECmTs, the so-called PBFDO or n-PBDF, which presents outstanding electrical conductivity and stability.[2] n-PBDF OECmTs exhibit non-volatile memory and long-term synaptic functions when H⁺ are used as migrating ions.[3] The effectiveness of the discussed strategies will be validated through simulations with deep neural network (DNN) models for handwritten digit recognition.[4] This talk aims to demonstrate the importance of understanding the synaptic mechanism to design strategies that can boost performance in neuromorphic devices.