Advanced machine learning for predictive 2D and 3D abrasion modeling in sediment bypass tunnels

Sediment bypass tunnels (SBTs) are widely used to alleviate reservoir sedimentation; however, they are susceptible to hydro-abrasive erosion, which can compromise their longevity. Predicting spatial abrasion is challenging due to the complexity of flow–sediment interactions and the scarcity of data. In this work, we synthesize recent advances in using machine learning (ML), specifically the XGBoost algorithm, to map abrasion patterns on SBT floors in both two-dimensional (surface damage/no damage) and three-dimensional (damage depth) contexts. Utilizing high-resolution laser-scan data from the Koshibu SBT in Japan, we extract geometric (e.g., distances along the tunnel and curvature) and hydraulic (e.g., flow velocity, discharge) features for each surface point. For 2D mapping, points were classified as “abraded” or “not abraded” based on a 2 cm depth threshold; for 3D mapping, the actual damage depth was predicted. XGBoost was trained on 50% of the approximately 1.16-million-point dataset and tested on the remainder. The 2D classification model achieved an overall accuracy of 86.4% (with a sensitivity of 91.9% and a specificity of 89.6%), successfully identifying four characteristic abrasion patterns in both straight and curved sections. The 3D regression model achieved a correlation coefficient R of 0.861 and Nash–Sutcliffe Efficiency of 0.748 between the predicted and measured depths (exceeding 0.90 and 0.85 in some sections). These results demonstrate that ML models can effectively capture complex abrasion behavior, though small-scale “wave-like” damage patterns remain challenging. This integrated approach highlights the promise of modern ML techniques for predictive tunnel maintenance and sediment management in hydropower infrastructure.