Oscillatory computational networks based on coupled VO₂ oscillators via tunable thermal triggering

Abstract

Deep neural networks excel at tasks like classification and speech recognition but demand rising power for large datasets, motivating energy-efficient architectures. Networks of coupled oscillators are promising if their interactions are tunable - typically via extra electronics. This thesis introduces a compact thermal-trigger element made from VO2 to control synchronization of closely spaced VO2 oscillators. Thermal coupling lowers net energy compared with independent oscillation. With active tuning we experimentally implement AND, NAND and NOR gates and spiking-neuron firing patterns. Large-scale VO2 spiking networks achieve 90% accuracy on MNIST, demonstrating a novel route to computation with thermally coupled oscillators.