An online, continuous learning framework for training surrogates of ocean models

Artificial intelligence and machine learning methods are rapidly becoming an additional tool for scientific research. The intersection of these data-driven techniques with traditional, numerically based models largely centers around training compact models for sub-gridscale parameterizations or full-model surrogates. Within the oceanographic literature in particular, most of these applications train on static data, i.e. model output from a previously run simulation. For practical reasons, these data are must be coarsened in time and/or space and thus may alias some of the fundamental dynamics of the underlying system. In this presentation, we introduce a completely in-memory framework based on the open-source library SmartSim that addresses problems with data volume and model training. Using this framework we demonstrate the practicality of training a neural network surrogate using data streamed from a 1/4-degree global, MOM6 simulation. By intelligently sampling the data and continuously training a neural network, we demonstrate how this ability to train on timestep-level data can lead to a smaller, more compact network which can mimic the solution space with better fidelity. We conclude with other examples of hybrid AI/simulation techniques, used in other scientific domains, that also benefit from this framework.