Transfer learning of neural surrogates on multifidelity groundwater simulations

Computationally inexpensive surrogates of process-based models, such as deep neural networks, enable ensemble-based computations used in risk assessment, data assimilation, etc. However, generation of large datasets required to train a neural network can be as expensive as the ensemble simulations themselves. We ameliorate this challenge by using data from multifidelity (MF) groundwater simulations and transfer learning (TL) to reduce data generation costs while maintaining model accuracy. As a computational example, we train a deep convolutional neural network (CNN) to reconstruct permeability fields from saturation maps derived from a multiphase flow model. Starting with very low- and low-fidelity data generated on increasingly coarse meshes, we pretrain the CNN, followed by output-layer training and fine-tuning using only a limited number of high-fidelity samples. We demonstrate the surrogate’s robustness when interpreting low-quality inputs—such as interpolated maps or data affected by noise—which has strong implications for the applicability in practical hydrogeological scenarios. This multilevel MF-TL strategy achieves a favorable trade-off between computational efficiency and predictive accuracy, significantly outperforming high-fidelity-only approaches under the same computational budget. • Transfer learning (TL) improves both efficiency and accuracy in training CNN surrogates for inverse modeling, enabling high performance with limited high-fidelity data. • TL mitigates architectural mismatch by regularizing and adapting a network originally designed for a different task, enhancing model stability and generalization. • CNNs trained with multi-fidelity data exhibit improved robustness to real-world data sparsity and noise, making them more applicable in practical scenarios.