Abstract: This study addresses the limitations of conventional methods in groundwater contamination source identification, including inversion uncertainty, low computational efficiency, and restricted applicability. We propose a novel deep learning-based approach for pollution source recognition. A hydrogeological conceptual model was established, integrated with multiphase flow numerical simulations to develop contaminant transport models. This framework transforms traditional inversion challenges into deep learning forward-solving processes. Through orthogonal experimental design, we generated comprehensive input datasets spanning parameter spaces. A multilayer perceptron (MLP) model was constructed by coupling multiphase flow simulations with Bayesian hyperparameter optimization, forming an adaptive deep learning framework for groundwater pollution tracing. The developed MLP model achieves an average coefficient of determination (R²) of 0.79, with overall mean relative error controlled at 25.12%. All output features demonstrate recognition accuracy exceeding 70%.This research establishes a new technical pathway for contaminant source identification in complex groundwater systems, which can be quickly applied to different contaminated site conditions.

Key words: groundwater contamination, contamination plume source identification, numerical simulations, multilayer perceptron (MLP), deep learning