Prediction of thick plate properties of CLT and innovative panels through machine learning algorithms

Abstract

Determining the membrane, bending, and shear stiffness of cross-laminated timber (CLT) and innovative panels requires advanced finite element computations and homogenization techniques based on thin and thick plate theories. These computations may not be easily implemented by engineers. To address this issue, this paper implements and makes accessible a machine-learning model that directly predicts the linear elastic properties of CLT and innovative panels. First, a large database of plate stiffness moduli is built from finite element computations taking into account microstructural characteristics such as the layers’ thicknesses, board width, gaps width, longitudinal modulus of elasticity, and longitudinal and rolling shear stiffness moduli of timber. Second, three approximations are built and investigated: closed-form solutions, an artificial neural network trained on the database, and another artificial neural network that uses the closed-form solutions as prior knowledge before being trained on the data. The results demonstrate the superiority of artificial neural networks with prior knowledge and a very satisfying accuracy for engineering applications. The root mean square percentage error (RMSPE) values range from 0.23% to 3.01%, and the maximum absolute percentage error (MaxAPE) values range from 1.12% to 20%.