Utilizing agricultural wastes for fired bricks: A machine learning approach to compressive strength and water absorption predictions

The use of agricultural wastes in brick production is increasing due to their potential for sustainable construction and efficient waste utilization. Predicting the physical and mechanical properties of such bricks remains challenging because of complex interactions among process variables and waste materials. This study addresses this by developing predictive models using four machine learning (ML) algorithms, namely random forest regressor (RFR), extreme gradient boosting (XGBoost), artificial neural network (ANN), and ridge regression (RR), based on a dataset of 110 data points including bricks with agricultural wastes such as rice husk ash (RHA) and wheat husk (WH), along with the physical and processing parameters. The results indicate that all models show strong potential for predicting brick properties with optimized hyperparameters. RFR achieved the highest predictive performance (R2 = 0.879 for compressive strength, 0.901 for water absorption), followed by XGBoost and ANN, which showed moderate predictive ability but signs of overfitting; RR performed the least effectively. SHapley Additive exPlanations (SHAP), Partial Dependence Plots (PDP), and Individual Conditional Expectation (ICE) plots revealed that manufacturing parameters were the most influential features. Sensitivity analysis showed that soil content (RMSE↑ 7.90), firing temperature (RMSE↑ 5.40), and brick size (RMSE↑ 4.95) had the highest impact, whereas waste additives exhibited low sensitivity (RMSE↑ < 2.0), supporting their sustainable inclusion. This study introduces a holistic workflow integrating predictive modeling, interpretable ML tools, and sensitivity analysis, as a template for materials science, highlighting its potential to optimize waste-based fired bricks and provide a transferable methodology for sustainable construction applications.