Presentation Information
[P002A]Geology-Constrained Data Generation for Tunnel Construction Using Tabular GANs
○Yulin Xu1[Student presentation: Doctoral course], Yuna Nakazawa1, Natsuo Okada1, Yoko Ohtomo1, Youhei Kawamura1 (1. Hokkaido University)
Keywords:
Tunnel Construction,Conditional Tabular GANs,Synthetic Data Generation,Overbreak Prediction
Overbreak and blast-related uncertainties remain persistent challenges in tunnel construction, where excessive excavation can lead to increased costs, safety concerns, and construction delays. Accurate prediction of such behaviors is crucial for supporting tasks such as overbreak risk estimation and blasting parameter optimization. However, these predictions require large volumes of structured, high-quality geological data—such as rock strength, weathering degree, and fissure presence—that are costly and difficult to obtain in practice.Artificial intelligence (AI) has the potential to improve predictive capabilities in this domain, but its performance depends heavily on the availability of training data. This data scarcity limits the practical use of AI in tunnel engineering.To address this issue, we propose a geology-constrained data generation framework based on Conditional Tabular GANs (CTGAN). By embedding domain-specific knowledge—such as rock weathering grades, face conditions, and crack intensity—into the generative process, the framework produces synthetic data that are both statistically realistic and geotechnically valid. Logical constraints are enforced during training to ensure consistency and avoid implausible feature combinations.Applied to real tunnel datasets from Japan, the proposed model successfully replicates geological patterns and enables AI models to operate more reliably under limited data conditions. The generated samples support downstream applications such as overbreak prediction and blast analysis, providing a practical step toward broader AI adoption in tunnel construction.