<p>An explainable machine learning framework was developed to predict the total-stress cohesion and friction angle of Bhimal fiber-reinforced expansive black cotton soil under consolidated undrained conditions, using 216 laboratory observations. Nine input variables were considered: fiber characteristics, compaction state, moisture condition, and curing period. Benchmarks included linear regression and decision tree models, and ensemble models were evaluated using random forest, XGBoost, CatBoost, and LightGBM. Model performance was evaluated using a group train-test split and five-fold cross-validation. The ensemble models consistently outperformed the benchmarks. The independent test sets showed that XGBoost was the best predictor of cohesion (R² = 0.873) and that LightGBM was the best predictor of friction angle (R² = 0.994). Cross-validation showed that LightGBM was more stable for cohesion and CatBoost was more stable for friction angle. SHAP analysis identified fiber content as the dominant predictor of both shear strength parameters. The observed effect was attributed to tensile resistance, crack bridging and soil-fiber interlocking. The developed framework provided an interpretable approach for predicting the shear strength response of natural fiber-reinforced expansive soil under consolidated undrained conditions.</p>