Explainable Bayesian-Optimized Machine-Learning Model for Predicting the Unconfined Compressive Strength of MICP-Treated Sands

Abstract Microbially induced calcite precipitation (MICP) is an environmentally sustainable technique for industrial byproducts, traditional chemicals, and mechanical techniques. MICP establishes proper mineral bonding (CaCO 3 precipitation) between soil particles, ensuring long-term strength and durability in soil stabilization. Determining the unconfined compressive strength (UCS) of an MICP-treated sand specimen in the laboratory is time-consuming and labor-intensive. This investigation employs Bayesian-optimized machine-learning (ML) models to predict the UCS of 327 MICP-treated sand specimens. A SHapley Additive exPlanations (SHAP) analysis reveals calcite content as the dominant predictor of UCS (mean|SHAP| = 0.66), with a strong positive correlation, while biological parameters (urease activity, OD 600 ) and initial void ratio ( e 0 ) exhibit secondary nonlinear influences. Therefore, the gradational and MICP parameters are used as input variables. The ML models are trained on 5k-, 8k-, and 10k-fold configurations. The model analysis reveals that the model’s performance improves with k-fold. It is also observed that the Bayesian-optimized ML models exhibit curve-fitting issues, ranging from underfitting to overfitting. Still, the ensemble tree (ET_10k) predicts the UCS with a root mean square error of 0.2120 MPa and correlation of 0.9687.