A neural architecture search—multilayer perceptron—transformer approach for modeling hydraulic head fields in karst conduit—matrix systems

Karst aquifers are characterized by complex conduit–matrix structures and unclear hydraulic connectivity, making it difficult to accurately capture the spatial variability of hydraulic head fields. This poses significant challenges for regional water resource management and groundwater exploitation and protection. Traditional physical experiments are constrained by experimental conditions, scale effects, and observational limitations, making it difficult to fully elucidate flow processes within complex karst systems. Although computational fluid dynamics methods offer high accuracy, they are highly sensitive to boundary conditions and parameters, and involve substantial computational costs, limiting their applicability for multi-scenario and rapid predictions. To address these issues, this study proposes a neural architecture search (NAS)–multilayer perceptron (MLP)–transformer model to efficiently simulate the evolution of conduit–matrix hydraulic head fields in karst systems. The results show that the NAS–MLP–transformer model exhibits strong generalization capabilities across various scenarios, with R2 values up to 0.999 and root mean square error below 0.5 cm. Its predictive performance surpasses that of other machine learning models examined in this study, achieving high accuracy while significantly reducing computational time. This research provides an efficient modeling approach for hydrological processes in karst groundwater systems and offers valuable theoretical and practical support for refined regional water resource management.