NeuralWood

Wood’s natural variations, such as fiber orientation and growth rings, lead to nonlinear material behavior. In construction, this complexity is usually simplified by grading timber into narrow categories for standard models (e.g., FEM). While effective for structural safety, this approach often wastes resources, particularly in laminated timber like self-shaping bilayers, where oversized sections are used to compensate for uncertainty.

Our project seeks to improve material efficiency by developing a machine learning–based model that embraces, rather than ignores, wood’s natural variability. Using physics-informed neural networks (PINNs), artificial neural networks guided by physical principles, we can train accurate models with relatively small datasets. Coupled with inverse design optimization (IDO), this approach enables selecting the best timber boards based on their unique material features to meet performance goals with minimal waste.

We will (1) create a dataset of bilayer specimens using digital wood grading, 3D scanning, and computer vision; (2) develop PINN models to predict hygroscopic and mechanical behavior; (3) integrate PINN into IDO to design lightweight bilayer structures; and (4) fabricate and evaluate the results. By aligning computational design with natural variation, this research advances sustainable use of wood through smarter, data-driven material allocation.

ICD Institute for Computational Design and Construction, Department for Computing in Architecture,  University of Stuttgart

Zuardin Akbar, Laura Marsillo, Prof. Dr. Thomas Wortmann

MPA Material Testing Institute, University of Stuttgart

Dr.-Ing. Serena Gambarelli

Microtec Srl
Ante-Holz GmbH

Forschungsinitiative Zukunft Bau. Projektnummer:  10.08.18.7-24.40 

Akbar, Z., Gambarelli, S., & Wortmann, T. (2026). A machine learning based model for self-shaping wood bilayers: leveraging local natural variations for bending prediction. Advanced Engineering Informatics, 70, 104061. https://doi.org/10.1016/j.aei.2025.104061