Deep convolutional network based surrogate model for dynamic optimization of LNAPL remediation using multi-phase extraction

Abstract:

Multi-Phase Extraction (MPE) is a rapidly emerging, in-situ remediation technology for simultaneous extraction of vapor phase, dissolved phase, and non-aqueous phase contaminants from contaminated soil and groundwater. Coupling multi-phase simulation and optimization algorithms to improve the MPE efficiency for remediation of light non-aqueous phase liquid (LNAPL) contaminated sites is becoming popular. However, the simulation-optimization framework faces great challenges because of the multi-phase model complexity and significant computational burden. In this study, a dynamic optimization configuration is proposed to economically achieve the maximum LNAPL recovery effect, which is the well locations and pumping rates are dynamically optimized during the different remediation periods. A cheap-to-run surrogate model based on deep convolutional neural network (CNN) is developed to approximate and replace the multi-phase flow model to alleviate the computational burden. Output variables, including the pressure, NAPL saturation, and the mass fraction of VOCs in the aqueous phase from the multi-phase numerical simulator, are approximated from the trained CNN-based surrogate model, decreasing the computational time by several orders of magnitude. More importantly, this surrogate model with high-dimensional input-output allows for simultaneous consideration of the total extracted LNAPL mass and residual NAPL saturation distribution when evaluating the MPE efficiency, which is more practical and of a substantial advantage over previous works usually considering only the total extracted LNAPL mass. The proposed CNN-based dynamic optimization framework is evaluated using a 2-D multi-phase flow model of LNAPL remediation with three MPE remedy stages. Results indicate that the CNN-based dynamic optimization framework can effectively identify an optimal MPE design that substantially improves the remediation efficiency. The pattern-based and low-computational cost property provides possibilities for this method’s potential application in the dynamic remediation optimization of in-situ contaminated sites.