Inversion of groundwater contamination source based on a two-stage adaptive surrogate model-assisted trust region genetic algorithm framework

Abstract

The inversion of contamination source characteristics is of great importance to groundwater remediation and risk management. The surrogate model-based simulation-optimization method is the most commonly used groundwater contamination source inversion (GCSI) approach. However, the static surrogate model and the genetic algorithm (GA) suffer from inaccuracy and instability problems when the nonlinearity of the numerical model or the dimension of the variables to be identified increase. In this paper, a two-stage adaptive surrogate model-assisted trust region GA (TSASM-TRGA) framework was developed to improve the accuracy and stability of the complex high dimensional nonlinear GCSI. In this framework, the two-stage adaptive surrogate model was applied to balance the global and local accuracy of the surrogate model, and the trust region GA was adopted to accelerate the convergence and improve the stability of optimal solutions. By comparing the accuracy of the TSASM-TRGA, two-stage adaptive surrogate model-assisted GA (TSASM-GA), one-stage adaptive surrogate model-assisted GA (OSASM-GA), and static surrogate model-assisted GA (SSM-GA) frameworks, the results showed that the identified contamination source locations with TSASM-TRGA framework are more consistent with the actual locations, and the identified contamination source fluxes are closer to the actual ones. The mean relative error (MRE) between TSASM-TRGA framework and the actual values is 0.31%, which is much smaller than that for the other three frameworks. The inversion results with the TSASM-TRGA framework show a higher accuracy than those with the simulation-statistical method. The present study suggests that the proposed TSASM-TRGA framework provides a more effective approach to improve the accuracy and stability of GCSI.