Freeze-thaw model for compressive strength of flowable solidified soil based on machine learning

Freeze-thaw cycling is a critical factor triggering durability degradation of flowable solidified soil in cold-region engineering, and its damage evolution process is affected by multi-factor coupling with obvious nonlinear characteristics. Based on experimental compressive strength data of flowable solidified soil under freeze-thaw cycles, four influencing factors, namely solidifier dosage (SD), admixture dosage (AD), water-solid ratio (WS) and freeze-thaw cycle number (F-T), are selected as input variables. Four machine learning algorithms including Decision Tree (DT), BP Neural Network (BP), Support Vector Regression (SVR) and XGBoost are adopted to establish freeze-thaw damage prediction models for the compressive strength of flowable solidified soil. The SHAP interpretable framework is further employed for feature sensitivity analysis to quantify the contribution and nonlinear effect of each input factor on prediction results. The results reveal that the DT model suffers severe overfitting with drastic error fluctuations on the test set; BP and SVR achieve a high test‑set R 2 yet show noticeable prediction deviations; the ensemble tree model XGBoost presents well‑balanced fitting capacity and satisfactory overall prediction stability with low error dispersion. SHAP analysis indicates the feature importance ranking: F-T > SD > AD > WS. Freeze-thaw cycle number (F-T) acts as the dominant control factor, and high/low values of each factor impose distinct positive and negative effects on model outputs. The proposed modeling and interpretive analysis method can provide technical references for rapid prediction of mechanical indices of similar flowable solidified soil specimens and identification of primary influencing factors.