Enhancing inverse modeling in groundwater systems through machine learning: A comprehensive comparative study

Abstract

Tandem neural network architecture (TNNA) is a machine learning algorithm that has recently been proposed for estimating uncertain parameters with inverse mappings. However, its reliability has only been validated in limited research scenarios, and its advantages over conventional methods remain underexplored. This study systematically compares the performance of the TNNA algorithm to four traditional metaheuristic algorithms across three heterogeneity scenarios, each employing a specific inversion framework: (i) a surrogate model coupled with an optimization algorithm for cases with eight homogeneous parameter zones, (ii) Karhunen-Loève expansion (KLE)-based dimensionality reduction combined with a surrogate model and an optimization algorithm for a high-dimensional Gaussian random field, and (iii) generative machine-learning-based dimensionality reduction integrated with a surrogate model and an optimization algorithm for a high-dimensional non-Gaussian random field. Additionally, we evaluate algorithm performance under two different noise-level conditions (multiplicative Gaussian noise with standard deviations of 1 % and 10 %) for normalized hydraulic head and solute concentration data in the non-Gaussian random field scenario, which exhibits the most complex parameter characteristics. The results demonstrate that both the TNNA algorithm and the metaheuristic algorithms achieve inversion results that satisfy the convergence accuracy within these machine-learning-based inversion frameworks. Moreover, under the 10 % high-noise condition in the non-Gaussian random field, the inversion results remain robust when sufficient constraints are imposed. Compared to metaheuristic approaches, the TNNA method yields more reliable inversion results with significantly higher computational efficiency, highlighting the considerable advantages of machine learning in advancing groundwater system inversions.