From ESN to Physical Reservoirs for Space Weather Forecasting

Abstract Reliable prediction of space weather conditions, analogous in complexity and importance to terrestrial weather forecasting, is critical for protecting satellite infrastructure, communication systems, and power grids. Traditional numerical forecasting models, though powerful, struggle with high-dimensional chaotic dynamics, significant computational demands, and real-time applicability. Reservoir Computing (RC), a lightweight yet powerful framework utilizing recurrent neural dynamics, has recently emerged as an efficient alternative for forecasting complex, nonlinear phenomena. This study investigates and benchmarks five distinct reservoir computing architectures: Echo State Networks (ESNs), Nonlinear Vector Autoregressive (NVAR) reservoirs, the Lorenz system as a physical reservoir, and physically inspired mono- and bi-stable Duffing oscillator-based physical reservoir Computers (PRC). ESN and NVAR reservoir computers are popular and traditional reservoir computer algorithms, and Lorenz and Duffing oscillators are physically inspired reservoir computers. Each proposed model's predictive capability is rigorously evaluated using real-time solar wind magnetic field data, providing quantitative assessments and qualitative insights. Results reveal distinct advantages and limitations of each approach, highlighting that mono- and bi-stable Duffing oscillators can achieve performance close to traditional algorithm-based reservoirs in terms of accuracy and capturing multidimensional dependencies, while emphasizing the untapped potential of coupled oscillator-based PRC architectures.