Abstract: Large deformations of surrounding rock induced by tunnel excavation pose a common safety hazard in deep-buried soft rock tunnels, often leading to support structure failure, project delays, and even casualties. Traditional mechanical analysis methods rely on empirical assumptions and struggle to accurately predict deformation types and risk levels under complex geological conditions. This paper proposes an intelligent classification model based on a Plant Growth Simulation Algorithm-optimized Random Forest (PGSA-RF), which utilizes intelligent algorithms to globally optimize model parameters, significantly improving classification accuracy and efficiency. Validation results based on a global dataset of 130 tunnel engineering cases show that the PGSA-RF model achieves a classification accuracy of 91.2%, representing a 23.2 percentage point improvement over the baseline model, with superior stability and generalization capability. Feature importance analysis reveals that, within the scope of indicators collected in this study, the maximum horizontal stress, rock type, and integrity are the three dominant factors controlling the types of large deformations, providing a quantitative basis for on-site risk management. The model classifies individual cases in under 10 seconds, with a prediction error of no more than 8%, effectively addressing the prediction challenges of traditional methods under complex geological conditions. This offers a high-precision, high-efficiency intelligent analysis tool for risk assessment in geotechnical engineering.

Key words: Surrounding rock large deformation, Random forest, Plant growth algorithm, Hyperparameter optimization, Tunnel engineering