Graph attention neural network for subsurface stratigraphy on spatial and feature level using multiple-source sparse exploration data

• GNN-GAN framework consisting of five modules designed for 3D subsurface modelling. • Data fusion module facilitating the combination of multiple data sources. • Innovative node representation of exploration data in graph of subsurface. • GAN and resampling strategy to alleviate class imbalance in graph structured data. The ability to construct accurate subsurface profiles from limited site exploration data is essential for ensuring the safety and efficiency of geotechnical construction projects. Traditional statistical and probabilistic methods require significant computational resources, especially when dealing with three-dimensional subsurface modelling. Recent advances in graph neural networks (GNNs) for node classification tasks highlight their potential for subsurface profiling, where geological exploration data can be represented as labelled nodes. However, geological exploration data are challenging for GNN implementation due to resulting imbalanced graph data structures. To address this issue, this paper proposes a novel approach that integrates GNNs with generative adversarial networks (GANs) to enable simultaneous optimization of GNN and GAN using graph-structured exploration data and create subsurface model in a realistic and context-aware manner. By employing GANs and a resampling strategy, biases introduced by imbalanced data are mitigated through generating high-quality subsurface data that mirrors the inherent spatial relationships of geological structures. Ablation studies are fully conducted to examine the impact of different model configurations on prediction performance. The proposed approach is validated using exploration data from Suzhou metroline project and benchmarked on an open-source dataset, which proves to be superior to existing methods in terms of high accuracy and can be generalized across different geological domains. The quality of the generated data relies on the training set and biased real-world data may lead to unrealistic synthetic data generation. This can be potentially optimized through combining geophysical data to provide richer understanding of the subsurface.