Dresden 2026 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

DY: Fachverband Dynamik und Statistische Physik

DY 6: Machine Learning in Dynamics and Statistical Physics I

DY 6.4: Vortrag

Montag, 9. März 2026, 10:15–10:30, HÜL/S186

Learning single and multiple chaotic systems with minimal reservoir computers — •Francesco Martinuzzi and Holger Kantz — Max Planck Institute for the Physics of Complex Systems

Chaotic dynamics are present in a multitude of natural and engineered systems. Recently, chaos has been modeled using machine learning (ML) methods thanks to their ability to infer underlying governing equations without directly accessing them. Among ML models, echo state networks (ESNs) have been widely investigated because of their simple construction and efficient training. However, ESNs typically rely on randomly initialized reservoirs whose stochastic connectivity makes them difficult to interpret and tune. To what extent are random and complex reservoir topologies actually necessary for learning chaotic dynamics with ESNs? We show that deterministic constructions of the reservoir matrix outperform random initializations for the reconstruction of chaotic attractors. By testing ten distinct deterministic topologies against random reservoirs on over 90 different attractors, our results demonstrate consistently better performance for deterministic reservoirs. Furthermore, we show how the same deterministic reservoir topologies can be leveraged to learn multiple chaotic systems with a single reservoir computer, thereby showcasing multifunctionality.

Keywords: Reservoir computing; Chaotic systems; Chaos reconstruction; Minimal reservoir computers