Solving Complex Optimization Problems with Neuromorphic VO₂-based Oscillators

Abstract

Phase-encoded oscillating neural networks (ONNs) provide significant advantages over traditional von Neumann architectures built with metal-oxide-semiconductor (CMOS) technologies for solving complex optimization problems. These networks promise ultralow power consumption and rapid computational performance through a neuromorphic approach. They can be realized with phase-transition materials like vanadium dioxide (VO₂), which has drawn growing interest in the phase-change materials research community due to its remarkable low-power electrical and optical switching properties, near-room-temperature operation (68 °C), scalability, high endurance, high-frequency performance, and CMOS compatibility. Integrating VO₂ into large networks of oscillators on a silicon platform poses challenges related to stabilizing the correct oxidation state and fabricating structures with predictable, low-variability electrical behavior. This work addresses these challenges by studying the effects of various annealing parameters applied through slow thermal annealing, flash annealing, and rapid thermal annealing on the formation of VO₂ polycrystalline structures. An optimal substrate stack configuration, employing a hafnium oxide (HfO₂) interlayer between the silicon substrate and the vanadium oxide layer, is proposed to minimize device variability. Material and electrical characterizations are performed to assess device quality, and a step-by-step recipe for building reproducible VO₂ oscillators is presented. Up to nine of these oscillators are contacted simultaneously to create an ONN, solving nondeterministic polynomial time complexity problems, including Graph Coloring, Max-cut, and Max-3SAT. Using sub-harmonic injection locking techniques, the solution space is binarized, achieving optimal solutions within 25 oscillation cycles. Additionally, this study explores the use of VO₂ to develop brain-inspired FitzHugh-Nagumo neurons. A self-coupling synapse enables the generation of complex firing patterns, such as mixed-mode oscillations, promising for time-dependent computing problems. Compared to Kuramoto-based relaxation oscillators, FitzHugh-Nagumo oscillators demonstrate superior stability and circuit simplicity, excelling in pattern recognition applications. We demonstrate power efficiency and scalability surpassing current commercial technologies, paving the way for large-scale implementation of VO₂ ONNs.