Integrating AI and sustainable materials: machine learning approaches to wood structural behavior
¹Universidad de Medellín, Faculty of Engineering, Department of Computer Science,
Carrera 84 #30-65, CO 050026 Medellín, Colombia
²Universidad de Medellín, Faculty of Engineering, Department of Civil Engineering,
Carrera 84 #30-65, CO 050026 Medellín, Colombia
³Universidad EAFIT, Faculty of Engineering, Department of Computer Science, Carrera 49 N° 7 Sur - 50, CO 050022 Medellín, Colombia
*magonzalez@udemedellin.edu.co
Abstract:
Wood is a potential construction material that provides a renewable source for this crucial task compared to other classical materials, such as steel or concrete, with high carbon fingerprinting levels. This suitable material minimizes energy use and adds more sustainability to ecological consciousness. Tree planting promotes the balance of the carbon dioxide ecosystem and captures and stores greenhouse gas emissions. Wood also has peculiar characteristics in terms of its structural strength and thermal insulation, optimizing energy consumption by reducing the need for cooling or heating needs. To use this material in construction, it is mandatory to study the resistance parameters like compressive, tensile, and shear strengths, enabling it for great-span structural projects. The traditional modeling strategies used for characterizing stress-strain performances usually simplify the assumptions, overpassing the complex mechanical behavior of the wood under different physical conditions. Nonetheless, previous analyses have shown that the traditional models may exhibit significant deviations from the actual resistance parameters since they can be limited in predicting non-linear and anisotropic properties inherent in wood. To address these limitations, this study proposes using machine-learning-based regressors to predict the mechanical properties of wood. Notably, we propose Multiple Linear Regression models to preserve the model’s interpretability while preserving the ability to model the linear properties in the studied scenarios. Furthermore, we use metaheuristic models based on deep learning and ensemble methods to increase the goodness of fit of the predictions. We used an experimental campaign with a widespread type of wood characterization of different parameters under tension parallel to the grain, compression parallel and perpendicular to the grain, and shear conditions. The results showed a lower root mean square error (RMSE) and a higher determination index (R2). Preliminary results demonstrated the ability of machine-learning-based modeling to obtain more accurate and reliable mechanical behavior of renewable construction materials like wood.
Key words:
artificial neural networks, machine learning, random forest, sustainable materials