SHAP based interpretable machine learning for predicting pressuremeter modulus from common geotechnical tests

Pressuremeter modulus is a fundamental deformation parameter for foundation design, but its routine use is limited by the specialized testing it requires. This study developed an interpretable machine learning framework to predict pressuremeter modulus from standard geotechnical investigation data. The approach addresses a persistent gap in engineering practice: SPT measurements are abundant, but pressuremeter results are scarce. Six ensemble algorithms were evaluated using rigorous cross-validation on 215 paired measurements from comprehensive site investigations. CatBoost achieved the best predictive performance (R 2 =0.888, RMSE=8.10 MPa) after systematic hyperparameter optimization. Since ensemble models are inherently difficult to interpret, SHAP (SHapley Additive exPlanations) analysis was applied to quantify how individual features contribute to predictions. SPT N-value emerged as the dominant predictor, accounting for 39.7% of feature importance, followed by unit weight, moisture content, and test depth. The key innovation was extracting mathematical coefficients directly from SHAP marginal effects using linear regression on dependence relationships, thereby converting model insights into an implementable engineering equation. The resulting four-parameter transparent equation retained 86.8% of the ensemble’s accuracy (R 2 = 0.771, RMSE = 11.58 MPa) and can be implemented in spreadsheets without specialized software. Performance compared favorably with traditional empirical SPT-based correlations reported in the literature. While conventional multiple linear regression achieved slightly higher accuracy (R 2 =0.797), the SHAP-derived equation offers greater interpretability, which offsets this modest difference in practical engineering applications. The proposed framework provides practicing engineers with an interpretable tool for preliminary modulus estimation during routine investigations and offers a transferable methodological basis for converting complex machine learning models into transparent engineering tools for other geotechnical parameters.