A machine learning based model for self-shaping wood bilayers: leveraging local natural variations for bending prediction

Self-shaping wood bilayers exhibit uncertainties in bending behaviour due to natural material variations such as growth rings and grain patterns that influence properties such as differential swelling coefficients and stiffness. While previous physics-based models use simplified assumptions about these variations, our method uses machine learning to capture spatially varying local features from each sample. This data-driven method learns complex, non-linear relationships of these features with bilayer curvature. Our framework encompasses bilayer production, data acquisition and preparation utilising computer vision, as well as the development of machine learning models. We analysed 64 sycamore maple ( Acer pseudoplatanus L. ) bilayers, yielding 1,216 data points on spatially local natural variations that correlate with their curvatures. Random Forest proved to be the most accurate, achieving a coefficient of determination (R 2 ) above 0.9 during training and over 0.8 on unseen test data. This framework demonstrates the integration of fine-grained material data into predictive models for material-driven wood design, adaptable to different species and data-driven methods.