Efficient Approaches for Online Training of Gaussian Process Models in Multi-objective Coastal Groundwater Management. Presented at Early Career Hydrogeologists’ Conference - University of Birmingham, UK (30 November 2023)

In the realm of coastal groundwater management, data-driven surrogate models present an appealing option to replace intricate groundwater simulators, offering insights into aquifer responses to pumping with significantly reduced computational costs. When employing simulation-optimization for management, the development of an "online" trained surrogate entails continuous updates at each iteration. This involves integrating newly identified optimal pumping patterns into the training set, allowing the derived solutions to gradually converge towards those obtained with full-scale simulators. However, this approach can be computationally demanding, especially in multi-objective management scenarios with multiple non-dominated solutions. This study aims to investigate effective sampling strategies for the online training of surrogates. To achieve this, a two-objective pumping optimization problem is initially formulated based on observed hydrogeological conditions in San Salvador Island, Bahamas. Gaussian Process (GP) techniques are utilized to construct model surrogates, and four online sampling strategies are proposed. For any given pumping scheme, GP models are employed to emulate management objectives and constraints while quantifying associated uncertainties. Through repeated stochastic simulations using these fast GP models, the study assesses the probability of Pareto-optimality for each pumping scheme. The performance evaluation of each training strategy involves determining the average probability of Pareto-optimality and assessing the correlation between predictions. The findings of this study highlight the most suitable online sampling strategy for GP models among the proposed alternatives.