Title: Hyperparameter Search using Genetic Algorithm for Surrogate Modeling of Geophysical Flows

Abstract: The computational models for geophysical flows are computationally very expensive to employ in multi-query tasks such as data assimilation, uncertainty quantification, and hence surrogate models sought to alleviate the computational burden associated with these full order models. Researchers have started applying machine learning algorithms, particularly neural networks, to build data-driven surrogate models for geophysical flows. The performance of the neural network highly relies upon its architecture design and selection of other hyperparameters. These neural networks are usually manually designed through trial and error to maximize their performance. This often requires domain knowledge of the neural network as well as the problems of interest. This limitation can be addressed by using an evolutionary algorithm to automatically design architecture and select optimal hyperparameters of the neural network. In this paper, we apply the genetic algorithm to effectively design the long short-term memory (LSTM) neural network to build the non-intrusive reduced order model of the sea-surface temperature field.