A metaheuristic–machine learning framework for modeling and improving the thermal behavior of bio-based wall panel systems in residential buildings

Abstract

Bio-based wall panel solutions are being promoted as sustainable alternatives to conventional wall construction; however, the thermal performance of these solutions is influenced by complex interplays between material properties, geometric configurations, and environmental conditions. This paper proposes an integrated machine learning and metaheuristic approach to predict and optimize the thermal, economic, and environmental performance of bio-based wall panel solutions for residential buildings. The heterogeneous dataset was developed by incorporating experimental data, validated simulations, and literature data to train various predictive models. The accuracy of the advanced nonlinear machine learning models was found to be high (R² up to 0.85), which was better than the accuracy of the linear machine learning model and proved the existence of interaction-driven thermal behavior. The results of the explainability analysis identified wall thickness as the primary factor influencing the performance of the wall panel solutions, followed by porosity and moisture-related attributes. The optimization results quantitatively determined the trade-offs between thermal transmittance, embodied carbon, and cost and proved that low U-value wall panel solutions can be developed with moderate embodied carbon and cost levels. The results of this research demonstrate that it is possible to attain high levels of thermal efficiency with relatively low embodied carbon levels, especially for well-designed plant fiber wall panel solutions.