A part of the final year project solution for low cost,sprase and explainable situations using LSTM,PINN,KAN-based surrogate models for two scale dynamics forecasting.
Two-scale dynamical systems involve processes that operate on two distinct time or spatial scales, typically a fast scale and a slow scale. the dynamics of the slow scale are sually described as the macroscopic dynamics
To test the validity of the surrogates two dynamical systems from the PFE were chosen:
For model 1 we can rely on the hoogenization to extract the macroscopic behavior while for model 5 we rely on the rolling average.
For sufficiently large and informative training data a basic LSTM network is used as a mapping
For sparse training data a ResNet based PINN network is used as a mapping
For explainability a simple KAN network is used as a mapping