Machine learning prediction of bacillus sphaericus bio-concrete strength: a comparative study of XGBoost, random forest, MLP, and ANN

Bacterially enhanced bio-concrete incorporating Bacillus sphaericus offers autonomous crack repair through microbially induced calcite precipitation, but the nonlinear dependence of its mechanical performance on mix design variables complicates reliable strength prediction. This study develops and compares four machine learning models - Extreme Gradient Boosting (XGBoost), Random Forest (RF), Multilayer Perceptron (MLP), and Artificial Neural Network (ANN) for predicting the compressive strength of Bacillus sphaericus bio-concrete at curing ages of 7, 14, and 28 days. A dataset of 120 experimental observations, characterised by five input features (curing age, bacterial concentration, water-cement ratio, nutrient concentration, and cement content), was partitioned into training and test subsets (80:20 stratified split) and evaluated under 5-fold cross-validation. XGBoost achieved the highest predictive accuracy (R² = 0.97, RMSE = 2.85 MPa, MAPE = 1.98%), followed by RF (R² = 0.95), MLP (R² = 0.93), and ANN (R² = 0.91). SHAP-based feature importance analysis identified curing age and bacterial dosage as the dominant strength predictors, with bacterial concentration exhibiting a saturation effect beyond 5 × 10⁶ CFU/mL. These findings offer a data-driven framework for optimising bio-concrete mix design and reducing experimental trial requirements in sustainable infrastructure applications.