Remote training of a reservoir computer via digital twins.

The increasing energy consumption required for information processing has become a significant challenge, leading to growing interest in optical and optoelectronic reservoir computing as a more efficient alternative. Trained reservoir computers are especially suited for low-energy applications near the edge. However, the computational cost of training the reservoir output weights, particularly due to matrix operations, adds potentially unwanted complexity to the architecture. To lift this restriction, we propose a remote training approach using digital twins-virtual models that replicate the behavior of a physical reservoir. In particular, unlike traditional training methods, we do not need to record the reservoir states experimentally for every new task. This allows the physical reservoir to be used continuously for inference without interruptions. We constructed two types of digital twins: a differential equation-based model and a deep neural network (DNN) model. Using the proposed remote training on real experimental data for the Santa-Fe laser time-series task confirmed that both models successfully captured the dynamics of the optoelectronic reservoir, allowing accurate predictions and the export of weights from the digital twin to the real world. The equation-based model achieved higher prediction accuracy than the DNN model, while the DNN model demonstrated greater robustness to variations in hyperparameters. These results demonstrate that digital twins can effectively enable the remote training of reservoir computing systems.