Predicting CPT-based shallow foundation design adequacy using machine learning

Accurate classification of shallow foundation design adequacy is essential for balancing safety and cost-efficiency in geotechnical engineering. This study develops a machine learning framework that combines Cone Penetration Test (CPT)-derived features with simulated spatial soil variability to classify foundation designs as under-designed, adequate, or over-designed. Feature engineering was performed to capture both soil heterogeneity and geotechnical behavior, while class imbalance was addressed using SMOTE and model reliability enhanced through hyperparameter tuning. Four classifiers—Random Forest (RF), XGBoost, Support Vector Machine (SVM), and Logistic Regression (LR)—were systematically evaluated. XGBoost and LR achieved the highest overall accuracy (0.95), with LR excelling in detecting the minority over-design class (F1 = 0.88, recall = 1.00). RF provided strong precision and robustness, while SVM offered high recall, particularly for the safety-critical under-design class. The framework was validated not only on synthetic Gaussian Random Field (GRF) data but also on a large field dataset of CPTs, confirming its practical applicability. The results demonstrate the value of domain-informed feature design and algorithm selection, highlighting LR as the most balanced model. The proposed framework provides a decision-support tool for integrating machine learning into foundation design, enhancing reliability under complex and variable soil conditions.