PNNL

Abstract

The immense computational cost of traditional weather and climate models has sparked the development of neural network based emulators or surrogate models. Because neural networks benefit from long records of training data, it is common to use datasets that are temporally subsampled. Here, we investigate how this often overlooked processing step affects the quality of an emulator’s predictions. We implement two architectures that stem from a class of machine learning called reservoir computing: a form of Nonlinear Vector Autoregression (NVAR) and an Echo State Network (ESN). Despite their simplicity, it is well documented that these architectures excel at predicting low dimensional chaotic dynamics. We are therefore motivated to test them in the more general setting of predicting high dimensional geophysical turbulence as resembled by a Surface Quasi-Geostrophic model. In all cases, we see that subsampling the training data consistently leads to a bias at small spatial scales, resembling numerical diffusion. Interestingly, the NVAR architecture becomes unstable when the temporal resolution is increased, indicating that the polynomial based interactions are insufficient at capturing the detailed nonlinearities present in turbulence. The ESN architecture is found to be more successful, suggesting a benefit to the more expensive but more general structure. We show that the spectral errors can be reduced to a minimum by including a penalty on the kinetic energy density spectrum during training, although the subsampling related errors persist. Future work is therefore warranted to understand how the temporal resolution of training data affects other neural network architectures.