<p>A homotopy-variational progressive search system incorporating a brain-storm parallel stochastic searching (BSPSS) algorithm was designed to precisely characterize groundwater dense nonaqueous phase liquid (DNAPL) contamination sources and simultaneously estimate contaminant transport parameters. The system was supported by two machine learning models: a novel stacked inverse machine learning (SIML) model for preliminary inversion and prior information correction, and a Bayesian hybrid-kernel machine learning (B-HKML) model for highly reliable surrogate modeling of numerical simulations. This configuration significantly reduced the computational cost of iterative likelihood evaluations. Results demonstrated that the B-HKML model successfully approximated the forward simulation input–output mapping, achieving a mean relative error (MRE) of 1.1784% in predicting contaminant concentrations. In contrast, the inverse input–output mapping of the numerical model was more challenging to approximate. Although the SIML model exhibited significantly better pattern recognition capability than traditional machine learning methods, it still yielded an MRE of 5.2817%. Nevertheless, the prior information and initial starting points generated by the SIML approach effectively constrained the search space and search paths, thereby improving both the efficiency and accuracy of the stochastic search system. The homotopy-variational search mechanism was shown to be effective in rationally partitioning the search space. Meanwhile, the BSPSS algorithm utilized swarm intelligence to perform fine-scale searches through cooperative chains, enabling progressive convergence toward a reasonable posterior distribution that accurately reflected the actual contamination source characteristics.</p>