Explainable machine learning driven predictive modeling of biochar-based cementitious composite

Fazal Hussain, Nancy Soliman,
Explainable machine learning driven predictive modeling of biochar-based cementitious composite,
Journal of Building Engineering,
Volume 113,
2025,
113942,
ISSN 2352-7102,
https://doi.org/10.1016/j.jobe.2025.113942.
(https://www.sciencedirect.com/science/article/pii/S2352710225021795)
Abstract: Biochar offers a promising strategy for mitigating CO2 emissions and potentially transforming concrete into a carbon sink. Recent studies highlight its positive impact on the mechanical and durability properties of cementitious composites. However, the variability in biochar properties and pyrolysis conditions increases the design parameters, making the design complex. To address these challenges, this study aims to develop an effective and explainable machine learning framework for predicting the compressive, flexural, and fracture energy properties of biochar-based cementitious composites. Machine learning algorithms with metaheuristic optimization were trained and tested using a dataset of 874 experimental points, which included mix proportions, biochar characteristics, pyrolysis conditions, and curing conditions. Subsequently, the model was validated with a new experimental dataset. The results demonstrate that Extreme Gradient Boosting (XGBoost) with Grid Search Optimization (GSO) accurately predicts the compressive strength, with a testing R2 of 0.96 and RMSE of 3.23. Similarly, Stochastic Process Regression (SPR) with GSO and Bayesian Optimization (BO) demonstrate the highest accuracy for predicting flexural strength and fracture energy, with R2 values of 0.99 and RMSE values of 0.53 and 0.04, respectively. Additionally, the Shapely additive explanation (SHAP) and Partial dependence plot (PDP) algorithms were employed to elucidate the machine learning models, providing insights into how input features influence predictions and aligning these findings with existing experimental studies. The curing age, biochar dosage, surface area, particle size, carbon content, water-to-cement ratio, pyrolysis temperature, and heating rate were identified as the most critical features influencing the mechanical properties of the biochar-based cementitious composite. Furthermore, experimental validation demonstrated that XGBoost-GSO and SPR-GSO predictions closely align with experimental outcomes, confirming their practical applicability for biochar-based cementitious composites. This research advances the development of sustainable construction materials by automating the design process, reducing labor and material costs, and enhancing overall efficiency.
Keywords: Biochar; Pyrolysis; Biochar cementitious composite; Machine learning; Explainable artificial intelligence; Feature analysis; Fracture properties; Carbon sequestration