A novel sensitivity-driven multistage feedforward inverse modeling approach for source identification of organic-contaminated aquifer

Abstract

In the present study, a sensitivity-driven multistage combined search system was established to improve the search capability and accuracy of inverse modeling for the simultaneous identification of DNAPL contamination sources and contaminant transport (CTR) parameters. The system was assisted by a biomimetic intelligence stochastic search process and a homotopy-variational deterministic search process. A Bayesian hybrid learning machine (BHLM) model, designed for accurate surrogate modeling of DNAPL transport numerical simulation, was also embedded in the combined inversion system, thereby reducing the computational burden of iterative likelihood evaluations. The results revealed that BHLM was highly successful in approximating the forward simulation input-output mapping, with a mean relative error (MRE) of 1.1784% for the prediction of contaminant concentrations. The sensitivity-driven whale optimization stochastic search algorithm (SD‑WOSSA) employed swarm intelligence to achieve a multi-chain parallel stochastic search process and incorporated a sensitivity‑driven multistage feedforward mechanism to enhance the differentiated search capability for source characteristics and CTR parameters. The SD-WOSSA approach provided refined prior information and high-quality starting points, which constrained the search space and search paths for the subsequent homotopy-variational deterministic search process. The homotopy-variational search mechanism was effective in reasonably partitioning the search space and transforming the global search problem into several linked local search stages, with the relative identification error remaining below 5%.