Breaking the mold of simulation-optimization: Direct forward machine learning methods for groundwater contaminant source identification

Groundwater Contaminant Source Identification (GCSI) is important for addressing environmental concerns. Currently, it is widely achieved using the Simulation/Optimization (S/O) method. However, the utilization of optimization techniques may cause high computation costs and parameter equifinality issues. We introduce two innovative GCSI methods, Direct Forward Machine Learning (DFML) and One-Hot Machine Learning (OHML), utilizing the classical Artificial Neuro Network (ANN) model in the machine learning field. Both new methods eliminate the need for an optimization algorithm in GCSI, thus reducing the construction effort and improving efficiency. The first method, DFML can directly estimate eight parameters, providing valuable insights into the contaminant location, release history, and aquifer properties. The second method, OHML can estimate the spatial probability distribution of the contaminant location through one-hot encoding, addressing uncertainties in the contaminant source location estimations realized by DFML. Evaluations demonstrate that both methods exhibit satisfying performances. DFML can estimate contaminant location and aquifer properties with high accuracy; and estimate the release history information with moderate accuracy. The OHML correctly assigns the higher contaminant probabilities to regions containing true contaminant locations. The combination of DFML and OHML offers a comprehensive framework. This study contributes valid methodologies to the GCSI field.