A Machine Learning Led Investigation Predicting the Thermos-mechanical Properties of Novel Waste-based Composite in Construction

• Published in: Waste and biomass valorization Vol. 15; no. 9; pp. 5445 - 5461
• Main Authors: Mahamat, Assia Aboubakar, Boukar, Moussa Mahamat, Leklou, Nordine, Obianyo, Ifeyinwa Ijeoma, Stanislas, Tido Tiwa, Bih, Numfor Linda, Ayeni, Olugbenga, Ibrahim, Nurudeen Mahmud, Savastano, Holmer
• Format: Journal Article
• Language: English
• Published: Dordrecht Springer Netherlands 2024; Springer Nature B.V
• Subjects: Building materials; Competitive materials; Composite materials; Compressive strength; Construction industry; Construction materials; Data collection; Datasets; Decision trees; Engineering; Environment; Environmental Engineering/Biotechnology; Flexural strength; Heat conductivity; Heat transfer; Industrial Pollution Prevention; Learning algorithms; Machine learning; Mechanical properties; Original Paper; Performance assessment; Predictions; Regression analysis; Renewable and Green Energy; Root-mean-square errors; Thermal conductivity; Thermodynamic properties; Waste Management/Waste Technology; Decision Tree; Eco-friendly Composite; Gradient Boost; Linear Regression; Random Forest
• ISSN: 1877-2641; 1877-265X
• DOI: 10.1007/s12649-024-02538-9

The study explores the potential of machine learning (ML) in predicting the thermal and mechanical properties of earth-based composites reinforced with natural Borassus fruit fiber. The limited availability of large datasets for accurate predictions is a challenge in material science research, which this study addresses. The authors collected data on thermal conductivity, compressive and flexural strength through experiments and employed four ML techniques suitable for small datasets: linear regression (LR), random forest (RF), decision tree regressor (DTR), and gradient boosting (GB). Evaluation metrics were used to assess the performance of the ML techniques. Linear regression emerged as the most efficient, exhibiting significantly lower error values compared to the others (e.g., RMSE of 0.066 for thermal conductivity, 0.119 for compressive strength, and 0.04 for flexural strength), followed by random forest and decision tree. However, gradient boosting showed relatively poor predictive accuracy. This study demonstrates the successful application of ML for predicting the properties of earth-based composites with limited data, which could significantly reduce the cost and time associated with developing new building materials and products. Manufacturers can gain a competitive edge by using ML to streamline material development, leading to lower costs, faster innovation, and the creation of more environmentally friendly building materials for a greener construction sector. Graphical Abstract