Multi-scale generative adversarial network for 2D subsurface reconstruction using multi-fidelity geological exploration data

Abstract

Accurate stratigraphic profiling from sparse site exploration data is crucial for geotechnical construction. Traditional methods, including manual delineation, geostatistical inference, and stochastic simulation, often face limitations such as oversimplification, high data demands, or dependence on expert assumptions, while probabilistic approaches require appropriate priors and domain expertise. This study proposes an intelligent computer-aided system for subsurface reconstruction using multi-fidelity geotechnical data using a Generative Adversarial Network (GAN). A multi-scale symmetric architecture is designed for unsupervised learning from a single image, incorporating a confidence-modulated noise map. Multi-fidelity data are fused through hybrid image representation to improve reconstruction accuracy. The GAN is trained on a benchmark geological profile and predicts on incomplete images with applied masks. Model performance is evaluated both qualitatively and quantitatively, with extensive ablation studies analyzing the impact of data fidelity and hyperparameters. Comparative results with state-of-the-art methods validate the effectiveness and efficiency of the proposed framework in integrating observed geological information into deep neural networks for realistic subsurface modeling and practical engineering applications.