Abstract: Saw-tooth joints serve as a crucial idealized model in rock mechanics research,where precise prediction of shear strength holds significant engineering implications.This study integrates swarm intelligence optimization technology with interpretable data-driven methods.Based on 50 sets of direct shear test data,four key parameters were selected as inputs and shear strength as the output to construct four machine learning base models and two swarm intelligence optimization composite models.Model performance was quantitatively evaluated through six metrics,with TreeSHAP method applied to analyze input parameter feature importance.It is found that the prediction performance of the group intelligent optimization combination mod el is significantly better than that of the traditional machine learning model.CSA-XGBoost is the optimal model,which can a chieve accurate prediction.The normal stress and joint inclination at joint surfaces were identified as core influencing parame ters,contributing over 58% cumulative feature importance.This study provides an efficient technical approach for quantitative assessment of shear strength in sawtooth jointed rock masses,offering valuable insights for intelligent prediction of similar rock mechanics parameters.Future work could enhance model generalization capabilities through sample supplementation and engi neering validation.

Key words: ensemble intelligent optimization technique,TreeSHAP method,CSA-XGBoost,shear strength