Biological plausibility and stochasticity in scalable VO2 active memristor neurons

Neuromorphic networks of artificial neurons and synapses can solve computational hard problems with energy efficiencies unattainable for von Neumann architectures. For image processing, silicon neuromorphic processors outperform graphic processing units (GPUs) in energy efficiency by a large margin, but they deliver much lower chip-scale throughput. The performance-efficiency dilemma for silicon processors may not be overcome by Moore’s law scaling of complementary metal-oxide-semiconductor (CMOS) field-effect transistors. Scalable and biomimetic active memristor neurons and passive memristor synapses form a self-sufficient basis for a transistorless neural network. However, previous demonstrations of memristor neurons only showed simple integrate-and-fire (I&F) behaviors and did not reveal the rich dynamics and computational complexity of biological neurons. Here we show that neurons built with nanoscale vanadium dioxide active memristors possess all three classes of excitability and most of the known biological neuronal dynamics, and are intrinsically stochastic. With the favorable size and power scaling, there is a path toward an all-memristor neuromorphic cortical computer.

💡 Research Summary

The paper presents a breakthrough in neuromorphic hardware by demonstrating that nanoscale vanadium dioxide (VO₂) active memristors can serve as biologically plausible artificial neurons with intrinsic stochasticity. Traditional silicon‑based neuromorphic processors excel in energy efficiency but suffer from limited throughput and poor scalability when trying to emulate the rich dynamics of real neurons. The authors address this gap by exploiting the Mott insulator‑to‑metal transition in VO₂, which yields a reversible “S‑shaped” negative differential resistance (NDR) region. When the circuit operating point is placed within this NDR region, the device becomes locally active, providing voltage gain analogous to the voltage‑dependent conductance of ion channels in biological membranes.

A two‑stage circuit is constructed using two VO₂ memristors biased in opposite polarities to mimic Na⁺ and K⁺ channels, each coupled to a membrane capacitor (C₁, C₂) and a load resistor (R_L1, R_L2). The system is described by four coupled first‑order differential equations for the normalized metallic radii (u₁, u₂) and the membrane potentials (V_Na, V_K), matching the dimensionality of the classic Hodgkin‑Huxley model. This architecture enables the neuron to exhibit a full spectrum of excitability classes (Class 1, 2, and 3) and a host of dynamical phenomena observed in real neurons.

Experimentally, three circuit topologies are explored: (1) a resistively coupled “tonic” neuron that produces continuous spiking and bursting under steady DC current; (2) a capacitively coupled “phasic” neuron that fires a single spike or a brief burst at stimulus onset (Class 3 excitability); and (3) a mixed‑mode neuron with parallel R‑C coupling that displays hybrid spiking patterns. Across these configurations, the authors experimentally verify 23 distinct neuronal behaviors, including tonic spiking, tonic bursting, phasic spiking, phasic bursting, sub‑threshold oscillations, integrator, resonator, bistability, inhibition‑induced spiking/bursting, rebound spikes/bursts, spike‑frequency adaptation, spike latency, threshold variability, depolarizing after‑potential, accommodation, excitation block, all‑or‑nothing firing, refractory period, and stochastic phase‑locked firing (skipping).

A key contribution is the demonstration of intrinsic stochasticity arising from thermal‑induced fluctuations in the VO₂ transition. This stochasticity manifests as variability in spike timing and enables hardware‑level implementation of probabilistic computations such as Bayesian inference, which are difficult to achieve with deterministic CMOS neurons.

From a fabrication standpoint, the VO₂ devices are electroform‑free, CMOS‑compatible nano‑crossbars with critical dimensions ranging from 50 nm to 600 nm, a device‑to‑device switching‑threshold variation of < 13 % (coefficient of variation), and endurance exceeding 26.6 million cycles. Power analysis shows static power consumption in the tens of nanowatts per square micrometer and dynamic power in the sub‑microwatt range, projecting a neuron density and energy efficiency competitive with biological tissue. Simulations suggest that scaling to cortical‑scale arrays would preserve these advantages, offering a path toward an all‑memristor neuromorphic cortical computer.

In summary, the work establishes VO₂ active memristors as a viable, scalable, and energy‑efficient building block for neuromorphic systems that faithfully replicate the complex, stochastic dynamics of biological neurons, opening new avenues for transistor‑less, high‑throughput artificial intelligence hardware.