Relieve the Demand for Labeled Data of Deep Learning Models for Hydraulic Conductivity Field Tasks in Groundwater Through Self‐Supervised Learning

Abstract Deep learning (DL) has shown great potential in solving groundwater problems but often requires large labeled data sets, which are expensive and time‐consuming to obtain. In this study, we introduce a self‐supervised learning approach based on a masked autoencoder (MAE)—an encoder‐decoder architecture that reconstructs randomly masked input patches—to reduce the dependence on labeled data in hydraulic conductivity field (HCF) modeling. The MAE is pretrained on a large set of unlabeled HCFs, enabling the encoder to learn transferable spatial features while the decoder reconstructs the original fields. We evaluate the pretrained encoder on three downstream tasks: macrodispersivity prediction, surrogate modeling of solute transport, and preferential channel identification. Across these tasks, using MAE‐extracted features allowed us to achieve comparable performance while reducing the amount of labeled data needed by approximately two‐thirds for macrodispersivity prediction, nearly half for solute transport modeling and over 60% for preferential channel identification. These results demonstrate that MAE‐based pretraining can effectively leverage unlabeled simulations to produce domain‐specific representations that accelerate model development, particularly in data‐scarce scenarios. This approach offers a scalable pathway for applying DL to a wider range of hydrogeological problems while lowering labeling costs.