• This study pioneered the development of a Borassus fruit fiber reinforced earthen matrix bio-composite with the fiber’s impact on the bio-composite’s crucial hygro-mechanical characteristics. • A novel primary dataset of experimentally build hygro-mechanical properties was established and expertly leveraged to train and validate a suite of ensemble learning models • A comprehensive comparative analysis of seven machine learning models, encompassing both ensemble and individual approaches, was conducted to accurately predict the bio-composite's hygro-mechanical behaviour. • Ensemble learning (EL) methodologies and linear regression exhibited superior predictive accuracy for both hygroscopic and mechanical properties compared to other evaluated models. This study uses machine learning (ML) to predict the effect of natural Borassus fruit fiber reinforcement on the hygro-mechanical properties of a bio-based earth-based bio-composite, enhancing sustainable construction. The study used experimental results of the hygro-mechanical properties to develop a primary dataset for training and testing ML models (validation of 80/20). Various models including stacking and voting ensemble learning (EL), linear regression (LR), support vector machine (SVM), gradient boosting (GBoost), XGBoost, and random forest (RF) were tested, with EL methods showing superior predictive performance for hygro-mechanical properties compared to individual models. Ensemble methods generally outperformed individual models. Specifically, gradient boosting and stacking EL achieved the highest R² of 99.2% for hygroscopic properties. For flexural strength, linear regression and stacking ensemble learning achieved perfect R² values of 100%. The developed models have the potential to optimize bio-composite design, predict material performance in various environments, and accelerate the development of sustainable construction materials. Further research is needed to enhance the performance of less effective models through hyperparameter tuning and exploring different kernel functions.
Mahamat et al. (Sun,) studied this question.