Integrated Evaluation of Freeze–Thaw-Induced Sandstone Degradation Using a Deep Learning–Assisted Modeling Framework

Freeze–thaw (FT) cycles gradually deteriorate sandstone and threaten the safety of underground structures. However, quantitative assessment remains difficult because available laboratory data are often incomplete and heterogeneous. This study develops a data-driven approach for evaluating FT-induced sandstone degradation by integrating multisource experimental data sets and a deep neural network. A total of 437 records are compiled, and a two-dimensional indicator tagging strategy combined with an extreme-value placeholder is used to handle missing information. The model processes intrinsic rock properties and environmental factors to predict six degradation indicators. Under a conventional train–test split, the model shows high predictive accuracy within the data range (coefficient of determination generally exceeding 0.90). A more stringent source-separated validation further indicates reduced but still interpretable predictive capability, suggesting that performance depends on the similarity of testing conditions to the training data. To examine practical applicability, the predicted indicators are introduced into a tunnel finite-element analysis through conventional parameter relationships, and the simulated responses follow observed trends. The study therefore provides a compiled data set and a baseline predictive workflow linking laboratory data and numerical analysis, while highlighting the need for improved missing-data treatment and physically constrained modeling in future work.