Deep learning based optimization under uncertainty for surfactant-enhanced DNAPL remediation in highly heterogeneous aquifers

Abstract

The optimization for surfactant-enhanced aquifer remediation (SEAR) of dense non-aqueous phase liquid (DNAPL)-contaminated aquifers is usually accompanied by uncertainties, which may arise from the characterization of complex aquifer heterogeneity and DNAPL source zone architecture (SZA) due to measurement sparsity. Optimization under uncertainty is computationally expensive as it involves an enormous number of model runs. The huge computational burden can be alleviated by utilizing a surrogate model for repeated model evaluations. However, most of the developed surrogates are often limited to low-dimensional optimization problems that only consider simplified aquifer heterogeneity. In this study, we developed a multi-objective simulation-optimization framework to optimize the SEAR schemes considering the characterization uncertainties from both highly heterogeneous aquifer permeability and complex SZA. A fast-to-run convolutional neural network (CNN)-based surrogate model was developed to approximate the high-dimensional and highly complex input-output mapping of the DNAPL multiphase flow simulation model. We first used the rejection sampling strategy to generate random realizations of permeability and SZA conditioning on their limited measurements and then formulated a multi-objective optimization under uncertainty problem based on these realizations. The developed 3-D CNN was trained and used as the surrogate for repeated model runs in optimization to identify the optimal SEAR schemes under uncertainty. A 3-D numerical experiment was used to test the performance of the CNN-based simulation-optimization framework. Comprehensive analysis on the obtained Pareto fronts demonstrates that the proposed framework can efficiently identify reliable Pareto-optimal solutions with a 99.8% speedup compared to the traditional optimization coupled with the forward model. Moreover, the optimization considering multiple realizations enables us to perform the risk assessment to locate the risk zone where the NAPL phase possibly exists after remediation, which provides useful information for decision-making.