Machine learning applications in predicting wood mechanical properties through non-destructive testing

This work was aimed at investigating, how non-destructively measured modulus of elasticity, specific gravity, moisture content, together with the information on wood type, could be used for prediction of various mechanical properties of wood, e.g. bending, compressive, shear, and tensile strength (parallel and perpendicular to the fiber direction), hardness and impact bending resistance. Unlike previous works, MOE is treated as an input variable, leveraging its sensitivity to the bulk material behavior for improved prediction accuracy. The approach extends beyond commonly studied species like spruce and poplar to a wider range of woods, providing a compre hensive assessment of model robustness and generalization. To that end, several machine learning approaches, namely, linear regression, partial least squares regression, random forest regression, support vector regression and support vector regression combined with a regressor chain, were exploited. The model’s accuracy was first evaluated using the coefficient of determination (R 2 ) score and root mean squared error(RMSE), followed by visual inspection of the shapes of the predicted and observed distribution functions for the said properties, as well as an assessment of the associated computational cost. Accordingly, support vector regression (SV R), and SV R com bined with a regressor chain, SV R(RC), capable of effectively learning non-linear relationships in the data and considering the potential correlations among different targets, showed superior performance in predicting the mechanical properties of wood. Focusing on the SVR-RC model, it was observed that the performance of the trained SV R(RC) model differed among the properties, being high for the properties parallel to the fiber direction; e.g. compressive strength and flexural strength as well as side hardness, assessed R 2 , obtained from 10 independent runs, around 0.95 in test dataset, while being lower for work to maximum load, indicated by R 2 of 0.756.