Simple, low-cost and accurate data-driven geophysical forecasting with learned kernels

Modelling geophysical processes as low-dimensional dynamical systems and regressing their vector field from data is a promising approach for learning emulators of such systems. We show that when the kernel these also learned (using flows, variant cross-validation), then resulting data-driven models are not only faster than equation-based but easier to train neural networks long short-term memory network. In addition, they more accurate predictive latter. When trained on observational data, example weekly averaged global sea-surface temperature, considerable gains observed by proposed technique in comparison with classical partial differential terms forecast computational cost accuracy. publicly available re-analysis daily temperature North American continent, we see significant improvements over baselines climatology persistence-based techniques. Although our experiments concern specific examples, general, results support viability methods (with kernels) interpretable computationally efficient forecasting large diversity processes.