Abstract: As a cost-effective groundwater remediation technology, the precise implementation of Monitored Natural Attenuation critically relies on reliable predictions of contamination plume evolution. To address the efficiency limitations of traditional numerical simulation methods, this study developed a general-purpose predictive modeling framework integrating three key components: orthogonal experimental design, Flopy-based automated modeling, and machine learning. Through scenario design and batch numerical simulations, we generated 200-31,200 multidimensional spatiotemporal datasets representing groundwater contamination plumes of varying scales. Leveraging these datasets, we systematically investigated the differential impacts of dataset sizes and the model training parameters on the performance of the static prediction (Multilayer Perceptron,MLP) and the dynamic spatiotemporal prediction (Spatiotemporal-MLP, ST-MLP) models. Our findings elucidate the coupling mechanism between the data scale and the model architecture: the static MLP achieved a contamination extent prediction accuracy of 0.97 with 1,000 training samples, whereas the dynamic ST-MLP required 23,400 samples to attain a spatiotemporal resolution accuracy of 0.991. Furthermore, optimizing temporal node sampling frequency (every 180 days) reduced the data generation costs by 40% for ST-MLP while maintaining accuracy above 0.98. This framework demonstrates the dual application values: (1) providing quantitative benchmarks for MNA feasibility assessments, and (2) enabling rapid spatiotemporal plume predictions to overcome the static and time-consuming limitations of conventional risk assessments, thereby establishing an efficient dynamic early-warning mechanism for groundwater contamination risks.

Key words: groundwater contamination, contaminant plume, spatiotemporal evolution, surrogate models, machine learning