Prediction and optimization of bending performance of bio-based I-joists with sandwich web using MNLR–GA modeling

Bio-based sandwich I-joists have attracted increasing attention as sustainable alternatives to conventional load-bearing wood products; however, the bending performance of hybrid lignocellulosic sandwich webs remains insufficiently understood due to the complex interactions among material composition and structural configuration. In this study, the bending behavior of bio-based I-shaped sandwich beams composed of laminated veneer lumber (LVL) flanges and multilayer sandwich web cores made from cattail–poplar particles were systematically investigated. The effects of key structural parameters, including web core density, number of kraft-paper-separated web layers, and jute fabric surface density, on the modulus of rupture (MOR) were evaluated through three-point bending tests. A multivariate nonlinear regression (MNLR) model was developed to explicitly describe the nonlinear dependence of bending strength on the studied parameters, and regularization was applied to improve model stability and limit overfitting. Model performance was assessed using R ² , RMSE , and MAE , and the validated MNLR model was subsequently coupled with a genetic algorithm (GA) to identify optimal web architectures that maximize bending performance. The results demonstrate that the number of web layers is the most influential parameter governing MOR, followed by core density, while the contribution of jute surface density is comparatively weaker. Scanning electron microscopy observations further revealed that increasing the number of web layers reduces core porosity and enhances interfacial bonding, thereby promoting more efficient stress transfer under bending loads. Overall, this integrated experimental and data-driven framework provides a reliable basis for predicting and optimizing the bending performance of hybrid lignocellulosic sandwich I-joists.