Physics-Informed Modeling for Wood Thermal Analysis and Prediction

Traditional architectural design often assumes wood is a homogeneous material, ignoring its complex, spatially varying thermal properties. While deep learning models can map wood RGB images to thermal responses, they often act as uninterpretable black boxes that prioritize statistical correlations over physical reality. This paper investigates whether integrating thermodynamic laws into deep learning architectures can improve the accuracy and interpretability of thermal predictions for heterogeneous wood samples.

The researchers developed two physics-informed frameworks to predict thermal responses from RGB images and testbed temperature maps:

• Physics-Informed Convolutional Neural Networks (PICNNs): These models incorporate the 2D steady-state heat transfer equation as a soft penalty term in the loss function, allowing the network to learn while being guided by physical laws.
• Physics-Integrated Convolutional Neural Networks (PInteCNNs): These models treat the heat equation as a hard constraint by embedding a differentiable numerical solver directly into the network architecture. The network predicts spatially varying physical parameters, which are then processed by the solver to generate the final thermal output.

To validate these methods, the team collected a multimodal dataset of three wood types (Poplar, Grandis Cross-Cut, and Grandis Radial-Cut), using an experimental setup that captures paired RGB and thermal images under controlled steady-state conditions.

The study demonstrates that embedding physical inductive biases significantly improves predictive performance compared to purely data-driven approaches. By enforcing thermodynamic constraints, the models successfully filter out non-physical measurement noise and provide interpretable physical parameters, such as thermal anisotropy. The PInteCNN approach, in particular, demonstrates that hard-coding physical solvers allows the model to maintain high accuracy while ensuring that the predicted thermal fields remain physically plausible, even when dealing with the inherent heterogeneity of wood.