Enhancing Inverse Modeling in Groundwater Systems through Machine Learning: A Comprehensive Comparative Study

Abstract. Machine learning has significantly improved inverse modeling for groundwater systems. One promising development is the tandem neural network architecture (TNNA), which integrates surrogate modeling and reverse mapping for efficient forward simulations and data assimilation. Although TNNA has shown success in groundwater inverse modeling, its application scenarios remain limited, and its advantages over conventional methods have not been fully explored. This paper aims to address these gaps by comparing the TNNA method with four conventional metaheuristic algorithms: Particle Swarm Optimization, Genetic Algorithm, Simulated Annealing, and Differential Evolution. Two synthetic solute transport numerical cases are designed, with aquifer parameters characterized by low- and high-dimensional scenarios, respectively. The surrogate model is constructed using a deep residual convolutional neural network (ResNet), selected based on a comparative evaluation against three other popular machine learning models. Inversion performance is evaluated based on the accuracy of calibrated hydraulic heads, solute concentrations, and parameter estimation errors. The results demonstrate that the TNNA algorithm yields more reliable inversion results and significantly reduces computational burden across both low- and high-dimensional cases, effectively balancing exploration and exploitation in global optimization. This study highlights the significant advantages of machine learning in advancing groundwater system inversions.