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

DY: Fachverband Dynamik und Statistische Physik

DY 42: Focus Session: Computing with Dynamical Systems: New Perspectives on Reservoirs and Applications I – Fundamentals

DY 42.6: Vortrag

Donnerstag, 21. März 2024, 11:15–11:30, BH-N 243

High Dimensional Hybrid Reservoir Computing — •Tamon Nakano¹, Sebastian Baur¹, and Christoph Räth^1,2 — ¹Institutfür KI-Sicherheit, Deutsches Zentrum für Luft-und Raumfahrt, Sankt Augustin/Ulm, Germany — ²Fakultät für Physik, Ludwig-Maximilians-Universität, Munich, Germany

Reservoir Computing (RC) is getting popularity as an alternative solution for complex dynamical systems, where physically derived models reach their limitation. RC is by default fully data-driven method and is expected to learn the underlying system in the dataset. However RC can't do so for a lack of data quantity, for example. The hybrid approach is now recognized as a powerful option for it. The idea is to combine a knowledge-based model as a support (e.g. an imperfect governing equation) to the fully data-driven method. This combination can be done at the input, output layer of RC or both of them (respectively called, input-, output-, full-hybrid). Some studies have been already done, for example, input- and full-hybrid by Pathak et al.(2018), output-hybrid by Doan et al.(2019). Duncan et al.(2023) compared the performance of the three approaches and showed the superiority of output-hybrid compared to the others. The prior studies above have developed the hybrid approach in lower dimensional problems (e.g. 3 dimension). In this work, we will extend the hybrid approach to higher dimensional systems. This will allow to treat highly nonlinear and time evolutionary systems with system knowledge, such as fluid dynamics simulations and time evolutionary phenomena captured in 2 dimensional images.

Keywords: Machine Learning; Reservoir Computing; Chaotic System; Hybrid Approach; Physics-Informed Approach