Prediction of Anchorage Support Parameters for Underground Caverns Using Neural Networks Enhanced by Variational Information Bottleneck

Accurate prediction of anchorage support parameters is important for the design of underground caverns. However, the limited availability of engineering cases makes it difficult to establish reliable data-driven prediction models. In this study, a neural-network-based framework was developed to predict anchorage support parameters using geological and engineering information from 121 underground cavern cases. Five input variables, including rock mass classification, excavation span, excavation height, uniaxial compressive strength, and maximum principal stress, were used to predict five support parameters. CNN and MLP models were established, and a variational information bottleneck (VIB) strategy was introduced to reduce redundant information and mitigate overfitting under small-sample conditions. The prediction performance of conventional and VIB-enhanced models was evaluated using mean relative error. The results show that the introduction of VIB improves the prediction accuracy of both CNN and MLP models, with the IB-CNN model achieving the best overall performance. An engineering application example further demonstrates the potential of the proposed method for preliminary anchorage parameter estimation. The proposed framework provides an auxiliary data-driven approach for underground cavern support design under limited data conditions.