Hardware-algorithm co-design in analog reservoir computing with nonlinearity of solution-processed 2D materials

Abstract

Reservoir computing, a recurrent neural network paradigm, shows potential in tracing chaotic dynamics in, e.g., motion tracking, spatiotemporal pattern recognition, and anomaly detection. However, the iterative nonlinear mapping required for reservoir activation poses challenges for digital computing. Realizing physical nonlinear systems from low-dimensional materials as the reservoir for performing analog nonlinear mapping emerges as a promising solution. Though promising, current advances remain largely at conceptual explorations via simulations, limited by the practical circuit design and fabrication challenges, and there has been a lack of hardware-algorithm co-design studies. In this work, we investigate hardware-algorithm co-design in analog reservoir activation with the nonlinearity derived from solution-processed two-dimensional (2D) materials. We show that the nonlinearity can be fitted as analog activation functions in implementing a reservoir computing model and, by co-design optimizations, the device parameterized model can achieve long-term synchronization and robust generalization in regression of chaotic systems, with resilience to noise. Given this performance, and the scalability of solution-processed 2D materials, the co-design scheme manifests the potential for the design and implementation of scalable, lightweight analog reservoir computing systems with solution-processed 2D materials for widespread applications in, e.g., IoTs, wearables, and robotics.