UC San Diego

Temperature and salinity profiles in the upper ocean are predicted using non-intrusive reduced-order modeling with deep learning. The data from Large Eddy Simulations, consisting of 8500 profiles of temperature and salinity fields, spanning over two weeks, are used to train the deep learning architecture. The entire two-week period is divided into phases: an active and a break phase. The deep learning architecture is trained in a two-step process. First, dimensionality reduction of the high-dimensional data is achieved using a Long Short-Term Memory(LSTM) based autoencoder. In the second step, another RNN-based architecture, the Gated Recurrent Unit (GRU), is used to evolve these reduced representations over time and then project them back into the high-dimensional space using the decoder, to recover the actual temperature and salinity fields. The primary goal of this work is to evaluate the prediction capability of the trained network under two scenarios: 1) using data from both phases for training, and 2) using only the active phase for training to predict the rapid evolution of temperature and salinity profiles between days 3 and 5. The secondary goal is to assess the model's long-term prediction capability by applying the same architecture used in scenario (2) of the primary goal to predict temperature and salinity profiles in the break phase and part of the active phase and highlight the importance of choosing the training data in predicting the physical phenomena.