Surrogate Modeling Coupled With Bayesian Active Learning Evaluated Using Information Theory Metrics for Computationally Expensive Sediment Transport Models

Complex process-based numerical sediment transport models suffer from computationally expensive forward runs, hindering various quantitative understandings of the investigating environment. For instance, inferencing physical parameters from field measurements, particularly those that cannot be directly measured in the field, requires many forward runs. This project focuses on constructing surrogate model(s) aided by active learning evaluated using information theory metrics during the training process, to emulate the full sediment transport model by speeding up the forward run while maintaining the quantitative relationships between the model inputs, underlying physical processes, and outputs, thus facilitating 1) the physical parameter posterior estimates, 2) realization of the model ensemble outputs, 3) and the model processes identification/selection. The trained surrogate model(s) can be further coupled with other inference methods such as simulation-based inference for even faster and likelihood-free posterior estimates.