Topologically Plastic Reservoir Computing with Self-rectifying Memristor Arrays

Abstract Predicting the evolution of high-dimensional chaotic systems, such as meteorological and climate dynamics, demands hardware-efficient frameworks capable of capturing complex spatio-temporal correlations. While reservoir computing (RC) offers a promising approach, conventional implementations are constrained by fixed network topologies that cannot adapt to the non-stationary, multi-scale dynamics inherent to real-world chaos. Here, we introduce a reconfigurable RC system with topological plasticity whose core functionality is intrinsically enabled by a novel self-rectifying memristor (SRM) array. The SRM device, featuring high rectification ratio (>106) and 128 distinguishable conductance states, exhibits a unique voltage-gated tri-modal operation: a low-voltage regime (0–2 V) for local nonlinear computation; a high-voltage regime (2.5–3 V) that induces stable resistive switching, enabling dynamic, hardware-level reconfiguration of global reservoir connectivity; and a negative-voltage regime (−2V–0) that blocks inter-conductance crosstalk. This tri-modal operation allows the physical hardware substrate itself—a 16 kb SRM array—to morph its interconnect topology in direct response to operational demands. When applied to the classical Lorenz system, our model achieves normalized mean squared error (NMSE) of 0.0613 with a core-array energy cost of 0.83 pJ/op. In meteorological dynamics forecasting, the system physically implements an adaptively tuned small‑world network reservoir, sustaining > 90% frame‑by‑frame accuracy over 30 days—a 7.6% improvement versus the fixed‑topology reservoir hardware. Our work establishes a new paradigm of adaptive neuromorphic computing driven by device physics, providing a transformative hardware solution for the energy-efficient and high-fidelity forecasting of complex environmental phenomena.