Abstract
Artificial intelligence and deep learning are currently reshaping numericalsimulation frameworks by introducing new modeling capabilities. Theseframeworks are extensively investigated in the context of model correction andparameterization where they demonstrate great potential and often outperformtraditional physical models. Most of these efforts in defining hybrid dynamicalsystems follow {offline} learning strategies in which the neuralparameterization (called here sub-model) is trained to output an idealcorrection. Yet, these hybrid models can face hard limitations when definingwhat should be a relevant sub-model response that would translate into a goodforecasting performance. End-to-end learning schemes, also referred to asonline learning, could address such a shortcoming by allowing the deep learningsub-models to train on historical data. However, defining end-to-end trainingschemes for the calibration of neural sub-models in hybrid systems requiresworking with an optimization problem that involves the solver of the physicalequations. Online learning methodologies thus require the numerical model to bedifferentiable, which is not the case for most modeling systems. To overcomethis difficulty and bypass the differentiability challenge of physical models,we present an efficient and practical online learning approach for hybridsystems. The method, called EGA for Euler Gradient Approximation, assumes anadditive neural correction to the physical model, and an explicit Eulerapproximation of the gradients. We demonstrate that the EGA converges to theexact gradients in the limit of infinitely small time steps. Numericalexperiments are performed on various case studies, including prototypicalocean-atmosphere dynamics. Results show significant improvements over offlinelearning, highlighting the potential of end-to-end online learning for hybridmodeling.