Interpretable TabPFN-Based prediction of unconfined compressive strength in chemically stabilized soft soils for transportation subgrade applications

Abstract Chemical stabilization is widely used to improve soft soils for road embankments and transportation subgrades in soft-ground regions. However, predicting their unconfined compressive strength (UCS) remains difficult because strength development depends on coupled effects of soil plasticity, compaction, binder composition, and curing age. This study develops an interpretable Tabular Prior-data Fitted Network (TabPFN)-based framework for preliminary UCS prediction of stabilized soft soils from the Mekong Delta, Vietnam. A dataset of 168 observations with seven predictors—liquid limit, plastic limit, maximum dry density, curing age, cement content, lime content, and fly ash content—was analyzed. TabPFN was applied using its default configuration and compared with four GridSearchCV-optimized models: support vector regression, k-nearest neighbors, Random Forest, and CatBoost. Performance was assessed using an 80/20 train–test split and five-fold cross-validation, while SHAP and partial dependence plots were used for interpretation. TabPFN achieved the strongest observed test performance among the five evaluated models, with R² = 0.925, RMSE = 80.25 kPa, MAE = 63.31 kPa, and MAPE = 7.50%, and a mean cross-validation R² of 0.877. Cement content and curing age were the dominant predictors, and the learned trends were consistent with established soil-stabilization mechanisms. The framework can support preliminary UCS estimation and candidate-mixture prioritization before laboratory verification. Its applicability remains limited to the investigated dataset and experimental domain, and broader use requires independent validation on larger and more diverse soil–binder systems.