Neural network prediction of bending strength and stiffness in western hemlock (Tsuga heterophylla Raf.)

Abstract

The stiffness and strength, modulus of elasticity (MOE) and modulus of rupture (MOR), as well as density, moisture content, microfibril angle and diffraction pattern coefficient of variation of azimuthal intensity profile (I_CV) was determined for 259 small clear specimens. These samples represent 38 old- and second-growth western hemlock (Tsuga heterophylla) trees harvested from several sites in coastal British Columbia, Canada. The data were analyzed by classic statistical regression techniques to reveal interrelations among the mechanical properties and the inherent wood properties. Simultaneously, the predictive power of artificial neural networks was evaluated with the same data set by employing several optimization techniques. Regression analysis of wood density and the flexural strength properties resulted in R² of 0.172 and 0.332 for MOE and MOR, respectively. The most efficient network model proved to be far superior demonstrating correlation coefficients with models for MOE ranging between 0.693 and 0.750, and the corresponding MOR models ranging between 0.438 and 0.561 in all testing phases. It is apparent that neural networks have the potential and capacity to self-train and become powerful adaptive systems that can predict the strength and stiffness of wood samples. The neural network analysis also revealed the importance level of each independent variable on both MOE and MOR properties.