Inversion of Groundwater DNAPL Pollution Source Based on DCNN Surrogate Model and Hybrid Homotopy-PSO with Feedback Iteration

Existing DNAPL groundwater source inversion approaches are confronted with prominent bottlenecks: shallow surrogate models often fail to capture strong nonlinear multiphase flow relationships, traditional heuristic optimizers suffer from premature convergence, and ill-posed equifinality further degrades inversion reliability, together with prohibitive computational costs from repeated multiphase numerical simulation. Taking a typical chemical-contaminated site in Northeast China as the research object, this study establishes a multiphase flow numerical model that fully reproduces the migration and transformation mechanisms of chlorobenzene-based DNAPLs after systematic generalization of the site’s geological and hydrogeological conditions. To drastically cut the computational burden incurred during iterative inversion, high-quality datasets are generated via parameter sensitivity analysis and Latin hypercube sampling, based on which a deep convolutional neural network (DCNN)-driven high-fidelity surrogate model is constructed and embedded into the optimization framework as an equality constraint. A separated nonlinear programming model is formulated to independently quantify pollution source characteristics and hydrogeological parameters, with the objective of minimizing the residual error between field-measured and numerically simulated contaminant concentrations. A hybrid homotopy-particle swarm optimization (HH-PSO) algorithm is further proposed to address the limitations of conventional optimizers, including strong dependence on initial guesses and susceptibility to local optima. On this basis, a closed-loop feedback iteration scheme is developed, where source identification and parameter calibration are implemented alternately with bidirectional constraints and progressive correction to continuously refine and stabilize inversion outputs. This work presents distinct innovations in the methodology, algorithm, and practical application of DNAPL groundwater source inversion. Results from synthetic benchmark cases and on-site field applications demonstrate that the DCNN surrogate model achieves far higher fitting accuracy than shallow learning approaches (e.g., Kriging and support vector regression), with the coefficient of determination R2 exceeding 0.99. After the feedback correction iteration procedure, the average relative error for retrieved source locations, release histories, and hydrogeological parameters drops to 3.72%, and the overall computational efficiency is elevated by approximately 99.84%. The integrated simulation–optimization inversion framework proposed in this work integrates monitoring signal denoising, multiphase numerical simulation, deep learning surrogate modeling, hybrid intelligent optimization, and feedback iterative correction. This integrated system effectively resolves core technical bottlenecks in DNAPL groundwater source inversion, such as nonlinear ill-posedness, equifinality induced by mutual interference between source terms and aquifer parameters, prohibitive computational costs of multiphase simulations, and premature convergence of traditional optimization algorithms. The established framework can serve as a robust theoretical foundation and technical tool for rapid, precise source tracing, pollution liability confirmation, and remediation design at complex contaminated sites.