Aiming at the problems of significant strength damage of sandstone in cold regions under freeze-thaw action and the complex research process of traditional test methods, a machine learning-based prediction model for the strength damage of freeze-thawed sandstone was proposed. By means of literature research, 117 sets of experimental data on strength damage of sandstone after freeze-thaw cycels were selected as the data set. The longitudinal wave velocity, dry density, saturated density, number of cycles and freeze-thaw period were selected as the input characteristics of the model, and the strength loss rate was used as the output characteristics. Six machine learning models such as decision tree, random forest, support vector machine, XGBoost, extreme random tree and CatBoost were used to predict the strength damage of sandstone after freeze-thaw cycles, and R², MAE, MSE and RMSE were used to evaluate the performance of the model. The results show that the CatBoost model has the best performance in all evaluation indicators. Its R² on the test set exceeds 0.93, and the MAE value, MSE value and RMSE value are 0.028, 0.002 and 0.039, respectively, showing strong generalization ability and good prediction effect. Further, the Bayesian optimization algorithm was used to optimize the hyperparameters of the CatBoost model, and the number of iterations, the maximum depth of the tree and L2 regularization were determined as the key parameters affecting the prediction performance of the model. Eigenvalue weight analysis shows that the order of importance of each variable to the model contribution is : number of cycles>freeze-thaw period>dry density>saturated density>longitudinal wave velocity. The research results can provide effective prediction tools and method references for rock mechanics engineering practice in cold regions.