There is broad interest in layered, two-dimensional (2D) semiconductors owing to their high optical absorption and ability to heterogeneously integrate with various substrates. In such layered semiconductors, mesoscopic imperfections, such as surface step edges, may affect charge carrier kinetics and device performance, but studying their effects on carrier transport with sufficient spatial and temporal resolution is a challenge. Using scanning ultrafast electron microscopy, we image the near-surface charge carrier kinetics with mu m/ps spatiotemporal resolution in single-crystal MoS2 samples with surface steps tens of nanometers in height. We find an approximately 3-fold increase in surface trap density near such surface steps compared to the uniform, flat surface regions. Despite the difference in trap state density, filling of the traps occurs over the same time scale (similar to 100 ps), but carrier recovery is much slower near the step compared to the uniform regions of the material. These results provide an approach to measure ultrafast dynamics and determine the required uniformity of layered materials depending on the speed limit constraints for specific applications.

Our growing computing needs, especially in applications that heavily rely on artificial intelligence (AI), motivate a search for new components that could substantially augment the performance of general-purpose digital computers. Beyond ON/OFF switching, new components with linear multistate analog resistive tuning, nonlinear volatile switching, spiking, oscillatory, stochastic and other complex functionalities could enable highly efficient neuromorphic computing schemes for AI information processing. Compared to the extreme multifunctionality of biological neurons, realizing all the above characteristics in a single, scalable analog component remains a grand challenge. Here we investigate electrochemical gating combined with localized thermal activation to program and switch a single, vertically integrated and dimensionally scaled electrothermal chemical random access memory (ETCRAM) with a channel and reservoir composed of phase-separated vanadium oxide. Closely related to electrochemical RAM (ECRAM), ETCRAM uses an integrated gate-heater electrode to overcome kinetic barriers that help retain states at ambient temperatures. In addition to synapse-like stable and programmable analog resistance states arising from redox-tunable phase coexistence, a single component exhibits neuron-like nonlinear conductance switching with a tunable threshold and self-driven dynamics owing to the thermally driven metal-insulator phase transition in vanadium dioxide. More broadly, we demonstrate that electrochemically stabilized phase coexistence could unlock analog electronics with novel functionality, stability, reconfigurability, and scalability.

Associative memories, especially those based on low-energy analog hardware, offer a promising primitive to handle the memory dominated era of large artificial intelligence (AI) models and in-sensor/edge intelligence. However, a critical but unsolved limitation of such systems is their iterative updates, which accumulate the native hardware noise, in addition to consuming high energy and latency. While emerging attention-equivalent softmax-based dense associative memories (SDAMs) reduce this cost through single-step recall, their use of dot-product similarity not only imposes hardware complexity due to the need for bipolar information but also suffers from obfuscation of closely related patterns during retrieval. Here, we introduce distance-aware in-memory SDAM (DIM-SDAM) implemented on foundry-manufactured 256 × 256 transposable memristor crossbars via algorithm-hardware co-optimization. The distance-based similarity operates with unipolar inputs (thus reducing hardware overhead by 50%), while preserving single-step convergence that minimizes accumulation of analog hardware noise. DIM-SDAM reconstructs Fashion-MNIST with near-perfect accuracy for corrupted inputs and further retrieves 100 EMNIST patterns with >95% accuracy, overcoming the linear storage capacity limit of Hopfield networks. DIM-SDAM also outperforms GPUs by 25× in energy efficiency, opening a pathway toward scalable, energy-efficient associative memories.

There is growing interest in correlated oxides that can switch between volatile resistance states when an electrical bias is applied, functioning as artificial neurons in neuromorphic computing systems. Most devices typically rely on first-order insulator-metal transitions (IMT). However, recent discoveries have shown that devices made of a second-order spin-transition material, such as LaCoO3 (LCO), can exhibit different or improved functionalities. Despite their significance, the microscopic details surrounding the formation of conductive channels have still been unreported. In this study, the spatiotemporal details of channel formation are revealed by using a combination of infrared (IR) and Raman microscopy. Comparison of LCO and materials such as VO2 reveals critical differences with important ramifications for computing. First, the findings indicate that LCO channels are narrower and more efficient than VO2, but they are also more sensitive to electric fields and disorder. Channels are found to repeatedly hop between different locations under steady-state oscillations, a behavior not previously reported. Additionally, memory effects at high bias are observed. The experiments, along with finite element simulations (FES), suggest that the spin transition in LCO may significantly influence channel nucleation, leading to an increased sensitivity of neuronal devices to disorder and electrode geometry. We discuss how the inherent stochasticity and memory effects could enable functionalities in neuromorphic computing.

Vanadium dioxide (VO2) is of interest for adaptive electronic applications such as neuromorphic neuristor devices and variable emissivity or tunable thermal control materials, thanks to its key property-a metal-insulator transition (MIT) at 68 degrees C that is accompanied by a dramatic change in electrical and optical properties. To improve performance in these roles, it is critical to develop approaches to engineer transport properties and the MIT behavior. While many documented techniques exist to modulate the MIT and film resistivities via lattice strain and chemical doping, less is known about the effects of ion irradiation on the intrinsic properties of VO2, despite the ability to control the spatial distribution of irradiation beams and the prevalence of high energy ion implantation in the semiconductor industry. The impact of irradiation of different acceleration energies on the responses of VO2 is of specific interest, as charged particle energy generally impacts both the resulting defect profile and corresponding transport behavior. Here, we demonstrate that 2 MeV He ions at equivalent calculated displacements per atom, in two different types of films, can create remarkable changes to the nature of charge transport in VO2, especially in the low-temperature insulating phase. Simulation of resulting changes in electrical conductivity reveals that He ion irradiation offers a strategy to increase both oscillation frequency and the signal transmission. These results provide insights into the intentional design of defect populations to modulate transport for neuromorphic VO2 devices.

A memristor chip can provide fast and energy-efficient spectrum analysis by using robust and denoised in-memory computing memory arrays.

In recent years, there has been an increase in the research interest in correlated oxides, partially due to their ability to transition between volatile resistance states when an electrical bias is applied. This property makes these materials suitable for artificial neuron design towards neuromorphic computing systems. While first-order insulator-metal transition (IMT) materials, such as VO 2 , have historically dominated the field, recent findings have revealed that a second-order spin transition, like the one present in LaCoO 3 (LCO), can also exhibit volatile switching behavior with new and improved functionalities. Despite the extensive research efforts directed towards device operation and the macroscopic effects observed, the microscopic mechanisms underlying the formation of conductive channels that enable device operation remain to be fully elucidated. We present spatiotemporal studies of channel formation in operando current driven LCO artificial neurons using a combination of infrared (IR) and Raman microscopies. These complementary experiments, which employ high spatial resolution Raman spectroscopy and high temporal resolution IR-imaging, are capable of elucidating the switching mechanisms, even in the context of fast self-oscillatory regimes. Integrating these results with current-voltage characteristics, we are able to establish a complete correlation between material local changes and device behavior. Furthermore, the interpretation of the spectroscopy information provides novel insights into the material's evolution. A comparison of LCO and materials such as VO 2 reveals qualitative differences. The findings indicate that LCO channels exhibit increased sensitivity to electric fields and disorder in comparison to VO 2 . It has been observed that channels repeatedly hop between distinct locations during steady-state oscillations, a phenomenon that has not been previously reported. We present experiments along finite element simulations, which were utilized to develop and validate a model that supports our hypothesis. The results of this study indicate that the spin transition in LCO may have a substantial impact on channel nucleation, enhancing the sensitivity of neuronal devices to disorder and electrode geometry. The crossover regime present in LCO, where high spin and low spin thermal fluctuations dominate the energy landscape, constitutes a fundamentally different scenario from a first order phase coexistence. During periods of current spiking, this sensitive energy landscape gives rise to a probabilistic competition between non-simultaneous filaments in the oscillatory negative differential resistance (NDR) regime or upon consecutive switching events. We present direct observation of the distribution of filaments in micron-scale devices and perform electrode engineering to tune the probability distribution using uniform and non-uniform electric field geometries. We show that fully volatile regimes are accompanied by short- and long-term memory effects at higher current bias. These semi-permanent changes in the material have the potential to be exploited for memory applications. These results highlight the role of stochasticity and material-specific details for exploiting spin-crossover materials in neuromorphic applications.

Flat regions of the neural network loss landscape have long been hypothesized to correlate with better generalization properties. A closely related but distinct problem is training models that are robust to internal perturbations to their weights, which may be an important need for future low-power hardware platforms. In this paper, we explore the usage of two methods, sharpness-aware minimization (SAM) and random-weight perturbation (RWP), to find minima robust to a variety of random corruptions to weights. We consider the problem from two angles: generalization (how do we reduce the noise-robust generalization gap) and optimization (how do we maximize performance from optimizers when subject to strong perturbations). First, we establish, both theoretically and empirically, that an over-regularized RWP training objective is optimal for noise-robust generalization. For small-magnitude noise, we find that SAM's adversarial objective further improves performance over any RWP configuration, but performs poorly for large-magnitude noise. We link the cause of this to a vanishing-gradient effect, caused by unevenness in the loss landscape, affecting both SAM and RWP. Lastly, we demonstrate that dynamically adjusting the perturbation strength to match the evolution of the loss landscape improves optimizing for these perturbed objectives.

BiVO4 photoanodes are promising for solar water splitting, with photogenerated electrons and holes preferentially reacting at top {010} and lateral {110} facets, respectively. However, the mechanisms driving this facet-dependent reactivity remain unclear. Here, we investigate facet-dependent photocurrent and material heterogeneity using correlative scanning photoelectrochemical microscopy (SPCM), electron beam induced current (EBIC) mapping, and mid-IR scattering scanning near-field optical microscopy (s-SNOM). SPCM measurements of 62 BiVO4 particles confirmed higher photocurrents at lateral {110} facets compared to top {010} facets, but unexpectedly revealed variations in photocurrent among lateral facets within the same particle. Variations in lateral facet surface termination could explain the intraparticle-level reactivity heterogeneity, consistent with theoretical predictions. Nano-FTIR spectroscopy and Raman microspectroscopy indicated significant materials chemistry heterogeneity within individual particles and facets that could be attributed to variations in lattice vibration distortions that enhance the overlap between Bi 6s and O 2p orbitals. The increased orbital overlap is significant as it potentially increases hole mobility in the valence band and potentially explains the lateral facet-dependent charge separation efficiency observed in photocurrent maps. Facet-dependent electrical and EBIC measurements showed no space charge regions at interfacet junctions or metal-BiVO4 contacts under vacuum, suggesting that photogenerated holes beneath top {010} facets are unlikely to transport to lateral {110} facets to drive water/sulfite oxidation. These findings indicate the potential influence of distinct bulk properties and surface termination chemistries across different particles and facets, highlighting the importance of carefully controlling defects and surface chemistry during sample growth to optimize photocatalytic performance.

Training neural networks has traditionally relied on backpropagation (BP), a gradient-based algorithm that, despite its widespread success, suffers from key limitations in both biological and hardware perspectives. These include backward error propagation by symmetric weights, non-local credit assignment, and frozen activity during backward passes. We propose Forward Target Propagation (FTP), a biologically plausible and computationally efficient alternative that replaces the backward pass with a second forward pass. FTP estimates layerwise targets using only feedforward computations, eliminating the need for symmetric feedback weights or learnable inverse functions, hence enabling modular and local learning. We evaluate FTP on fully connected networks, CNNs, and RNNs, demonstrating accuracies competitive with BP on MNIST, CIFAR10, and CIFAR100, as well as effective modeling of long-term dependencies in sequential tasks. Moreover, FTP outperforms BP under quantized low-precision and emerging hardware constraints while also demonstrating substantial efficiency gains over other biologically inspired methods such as target propagation variants and forward-only learning algorithms. With its minimal computational overhead, forward-only nature, and hardware compatibility, FTP provides a promising direction for energy-efficient on-device learning and neuromorphic computing.

Constrained optimization underlies crucial societal problems, for instance, stock trading and bandwidth allocation. However, it is often computationally hard, in that complexity grows exponentially with problem size. The big‐data era urgently demands low‐latency and low‐energy optimization at the edge, which cannot be handled by digital processors due to their non‐(parallel von Neumann architecture. Recent efforts using massively parallel hardware (e.g., memristor crossbars and quantum processors) employing annealing algorithms, while promising, have handled relatively easy and stable problems with sparse or binary representations, such as the max‐cut or traveling salesman problems. However, most real‐world applications embody three features, which are encoded in the knapsack problem, and cannot be handled by annealing algorithms—dense and nonbinary representations, with destabilizing self‐feedback. Herein, a post‐digital‐hardware‐friendly randomized competitive Ising‐inspired (RaCI) algorithm performing knapsack optimization, experimentally implemented on a foundry‐manufactured complementary metal‐oxide‐semiconductor‐integrated probabilistic analog memristor crossbar, is demonstrated. This solution outperforms digital and quantum approaches by over four orders of magnitude in energy efficiency.

Managing heat is a major challenge in modern silicon-based computers due to both large static and dynamic power dissipations. There is a growing perspective that heat can serve as an information carrier (instead of being treated as a useless by-product) in post-silicon devices, enabling new functions and on-chip energy recycling. In this review, we introduce how heat can be utilized as a degree of freedom in electronic devices, and how such devices may enable efficient computers.

Negative differential resistance (NDR) is a key electronic response enabling two‐terminal artificial neurons that can be achieved through different physical phenomena, including phase‐homogeneous current density and temperature (electro‐thermal) localizations and spatially‐localized metal‐insulator phase transitions (MITs). These two effects have been observed to occur sequentially in select electrically‐biased transition metal oxides. However, it is unknown why and under what conditions localizing behaviors precede MITs, particularly as a function of device length scale. To this end, the interplay between phase‐homogeneous electro‐thermal localizations and MITs is investigated in a 3D multiphysics simulation of a lateral thin film device, using the material properties of the prototype MIT material VO 2 . These findings demonstrate that the MIT is nucleated through dynamically localizing current density and temperature. A critical device width (≈0.7 µm in this study) is identified, below which both the electrically‐induced electro‐thermal and phase inhomogeneities cease to appear. It is demonstrated that the formation of spatial inhomogeneities directly relates to device dimensions, and demonstrate the decoupling of NDR from the MIT through device scaling relationships. These results provide insight into the material phenomena underlying the material's electrical responses, clarifying conditions under which spatial inhomogeneities form in electrically‐biased MIT materials.

BiVO 4 particles are attractive visible light-absorbing semiconductors to drive the water oxidation half reaction in a photoelectrochemical cell or for the overall water splitting reaction in a photocatalytic reactor. Ensemble-level and single particle-level measurements previously showed that photogenerated electrons and holes tend to accumulate and react at top {010} and lateral {110} facets, respectively. However, multiple possible surface terminations exist for these low-index facets, and it is unclear to what extent different surface facet terminations influence the overall current or reaction rate of a single BiVO 4 particle. Here we use correlative scanning photoelectrochemical microscopy (SPCM), facet-dependent electrical measurements, electron beam induced current (EBIC) mapping, and mid-IR scattering scanning near-field optical microscopy (s-SNOM) measurements to quantify the water and sulfite oxidation photocurrent among different lateral and top facets of 62 BiVO 4 particles. SPCM measurements confirmed that illuminating lateral {110} facets generates higher photocurrents for water and sulfite oxidation than top {010} facets. However, they also unexpectedly revealed that different lateral facets within the same particle produce varying photocurrents, despite electron microscopy images showing no clear structural differences between them. Facet-dependent electrical and EBIC measurements further indicated that no significant space charge regions form at inter-facet junctions or metal-BiVO 4 contacts under vacuum. This result is significant because it is highly unlikely that photogenerated holes generated in the particle interior beneath the top {010} facet transport to the lateral {110} facets and contribute to the water/sulfite oxidation reaction. Additionally, nano-FTIR spectroscopy revealed that materials chemistry heterogeneity exists not only between particles but also within individual facets of a single particle. These findings suggest the presence of distinct surface termination chemistries across different facets, consistent with theoretical predictions, and emphasize the need to control atomic-level surface termination during sample growth to enhance photocatalytic performance.

Artificial intelligence (AI) is pushing the limits of digital computing to such an extent that, if current trends were to continue, global energy consumption from computation alone would eclipse all other forms of energy within the next two decades. One promising approach to reduce energy consumption and to increase computational speed is in-memory analog computing. However, analog computing necessitates a fundamental rethinking of computation at the material level, where information is stored as continuously variable physical observables. This shift introduces challenges related to the precision, dynamic range, and reliability of analog devices - issues that have hindered the development of existing memory technology for use in analog computers. Here, we address these issues in the context of memory which stores information as resistance. Our approach utilizes an electrochemical cell to tune the bulk oxygen-vacancy concentration within a metal oxide film. Through leveraging the gate contact as both a heater and source of electrochemical currents, kinetic barriers are overcome to enable a dynamic range of nine decades of analog tunable resistance, more than 3,000 available states, and programming with voltages less than 2 V. Furthermore, we demonstrate deterministic write operations with high precision, current-voltage linearity across six decades, and programming speeds as fast as 15 ns. These characteristics pave the way toward low-power analog computers with potential to improve AI efficiency by orders of magnitude.

Phase coexistence in nanoscale electrochemical random-access memory (ECRAM) has recently been demonstrated to enable both information storage and extraordinary reconfigurability. These proof-of-principle demonstrations have left the mechanistic details of such a process unresolved. Particularly, the mechanisms that stabilize the multiple phases, and the underlying processes behind sustained memory retention, remain unclear, and are necessary to design such devices. Here we report microscale ECRAM devices composed of VOx, which enables us to directly probe the active region in an operando fashion using optical techniques. Using Raman mapping, we show the phase coexistence driven by the electrochemical injection of O vacancies to be spatially uniform (i.e., with no filaments). The stability was observed to be unusually long, with 1% loss over 14 years in ambient conditions. First-principles calculations of the oxygen vacancy formation energies in VOx further support the thermodynamic coexistence of multiple VOx phases and clarify the origin of the observed long-term retention in the ECRAM devices. Further, we demonstrate single devices that can be voltage programmed to exhibit synaptic, neuronal, and reconfigurable logic gate functionalities. Therefore, we not only uncover the phase coexistence mechanism that may help device design, but also demonstrate the circuit-level applications of reconfigurability.

Over the past decade, dendrites of neurons, which were previously thought to perform only information pooling and networking, have now been shown to express complex temporal dynamics, Boolean-like logic, arithmetic, signal discrimination, and edge detection. Mimicking this rich functionality could offer a powerful primitive for neuromorphic computing. Here, using Ovonic threshold switching in Sb-Te-doped GeSe, we demonstrate a single two-terminal component capable of self-sustained dynamics and universal Boolean logic in addition to XOR operations (which is traditionally thought to require a network of active components). We then employed logic-driven dynamics to detect and estimate the gradients of edges in images. The Ovonic switch exhibits properties of a half adder and a full adder in addition to discriminative logic accommodating inhibitory and excitatory signals. We show that this simple computational primitive offers a highly improved energy efficiency. As such, this work paves the path for potentially emulating dendrites for efficient postdigital neuromorphic computing.

We demonstrate the simulation of noise- dependent point cloud processing using a compact model of non-volatile memories (NVMs). We investigate how classification accuracy is affected by programming variations in NVMs, representing circuit noise, and distortions in point clouds, representing sensor noise. We employ a PointNet-based framework and explore how the inherent weight-sharing properties of PointNet can leverage NVM crossbars for energy-efficient processing. By benchmarking the performance of both NVM-based PointNet across various noise levels, we demonstrate the impact of different noise types on classification accuracy. Our findings show that while certain circuit and sensor noise degrade classification performance, our NVM-based PointNet achieves competitive results with reduced trainable parameters, providing a path toward neuromorphic 3D vision and computing with co-optimization of software and hardware. Our work highlights the potential for using NVM crossbars to efficiently handle noise-dependent processing in edge inference units.