BEOL-compatible 4F2 Single Crystalline Semiconductor Oscillator for Low-power and Large-scale Oscillatory Neural Network Hardware

Abstract

Oscillatory neural network (ONN), which is a novel neuromorphic system composed of oscillatory neurons coupled via synapses, is suitable for solving complex patterns. In this work, we demonstrated the feasible ONN hardware based on an InGaAs biristor, a single-crystal semiconductor exhibiting high reliability, uniformity, and repeatability. We first evaluated the oscillation characteristics of the InGaAs biristor-based oscillator (IBO). To enhance the operational efficiency of the ONN, we proposed a sub-harmonic injection locking (SHIL) method. This technique allows for precise control of the oscillatory behavior, resulting in improved performance and energy efficiency. In addition, we systematically demonstrated the coupled capacitors acting as synapses to control the weight in coupled IBOs. Finally, we perform simulations to validate the feasibility of our proposed device. Specifically, we tested the performance of a 3×5 ONN system in a −+pattem recognition task. We expect that the advantages of the IBO in terms of its cell size (4F2), low-temperature fabrication (< 100 °C), and high reliability will contribute to future advancements in 3D stackable ONN hardware systems.