Predicting Shallow Water Dynamics using Echo-State Networks with Transfer Learning
Abstract
We present how Echo-State Networks (ESN) can be used to predict the dynamics of the Shallow-Water Equations. Such ESN models represent a computationally efficient method as an initial-value solver. Therefore, they can be used for performing ensemble simulations and predicting evolution of averaged quantities or accessing uncertainty. Our approach also can be interpreted as equation learning where the same neural network is capable of reproducing dynamics of the initial-value problem for a large set of initial conditions. We also demonstrate that large-scale quantities such as momentum and averaged water height play an important role in predicting the dynamics. We introduce a transfer learning approach to quickly re-train the neural network and take such large-scale quantities into account.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFMNG13A..03T