A Data-Driven Workflow for Modelling Self-Shaping Wood Bilayer, Utilizing Natural Material Variations with Machine Vision and Machine Learning

Abstract

This paper develops a workflow to train machine learning (ML) models with a small dataset from physical samples to predict the curvatures of self-shaping wood bilayers based on local variations in the grain. In contrast to state-of-the-art predictive models, specifically 1.) a 2D Timoshenko model and 2.) a 3D numerical model with a rheological model, our method accounts for natural and unavoidable material variations. In this paper, we only focus on local grain variations as the main driver for curvatures in small-scale material samples. We extracted a feature matrix from grain images of active and passive layers as a Grey Level Co-Occurrence Matrix and used it as the input for our ML models. We also analysed the impact of grain variations on the feature matrix. We trained and tested several tree-based regression models with different features. The models achieved very accurate predictions for curvatures in each sample (R;0.9) and extend the range of parameters that is incalculable by a Timoshenko model. This research contributes to the material-efficient design of weather-responsive shape-changing wood structures by further leveraging the use of natural material features and explainable data-driven modelling and extends the topic in ML for material behaviour-driven design among the CAADRIA community.