Seeing the Heat: Synthesizing High-Resolution Wood Thermal Responses from Optical Imagery

Jingren Xie

cs.CVSep 22, 2026

Seeing the Heat: Synthesizing High-Resolution Wood Thermal Responses from Optical Imagery

Authors: Jingren Xie

Abstract

The thermal behavior of wood is a critical factor in advanced material assembly. However, pixel-level thermal analysis remains fundamentally constrained by the low resolution and noise inherent to infrared thermography. To address this, we introduce an end-to-end computational framework that synthesizes high-resolution thermal responses directly from wood RGB images. We first establish a core physical linkage: because spatial color variation in natural wood is driven by cellular anatomy, optical intensity serves as a reliable geometric proxy for the localized solid volume fraction. By leveraging this theoretical insight, we develop an automated finite-element-method data engine that maps pixel-level optical intensity to a 3D thermodynamic voxel grid, generating high-fidelity synthetic thermal responses. We find that 1) when the thermal conductivity along the thickness direction is uniform or linear, wood RGB images and their corresponding thermal responses exhibit extreme morphological similarities, and the lateral thermal diffusion acts as a low-pass filter that smooths out high-frequency details; 2) when the thermal conductivity along the thickness direction is random, such morphological similarities are destroyed, and wood's 3D structure dominantly governs its thermal response. We further utilize these synthetic thermal responses to supervise a neural surrogate model built upon the DINOv3 foundation model. Our results demonstrate that the neural surrogate model successfully internalizes the governing thermodynamic laws, thereby bypassing computationally expensive simulations and enabling high-resolution thermal inference. This methodology effectively bridges the semantic and thermodynamic domains, unlocking systematic, pixel-level analysis of fine-grained wood thermal responses. Project: https://zekifayes.github.io/seeheat

Figures & tables

Figure 1: Overview of the Visual-Thermal Computational Framework. The Data Engine acts as a physics-based solver, mapping a wood RGB image to a thermal conductivity tensor and solving Eq. 1 to produce a high-fidelity thermal response.

Figure 2: Stability of the system parameter

\eta

across multiscale spatial configurations. The parameter

\eta

is optimized using MAE across various

K

and

\sigma

for the Poplar, Grandis-CC, and Grandis-RC datasets. The absolute value of

\eta

remains tightly constrained within a narrow numerical band for each dataset, demonstrating high model stability.

Figure 3: Mesh convergence analysis for the 3D finite element thermal simulation. (Left) The calculated average top surface temperature

\bar{T}_{\text{top}}

quickly stabilizes across different discretizations

n_{z}

for varying high-resolution grids. (Right) The relative error of each configuration, evaluated against the

n_{z}=20

baseline, remains orders of magnitude below a conservative 0.5% error threshold, validating the numerical stability of the thin-plate regime.

Figure 4: Overview of the Wood Thermal Prediction Framework. It leverages this synthesized paired dataset to supervise a DINOv3-based encoder-decoder architecture, enabling direct prediction of the thermal responses from the wood RGB images.

	ViT-S			ViT-B			ViT-L
Decoders	MAE ( $\downarrow$ )	RMSE ( $\downarrow$ )	$\delta_{01}$ ( $\uparrow$ )	MAE ( $\downarrow$ )	RMSE ( $\downarrow$ )	$\delta_{01}$ ( $\uparrow$ )	MAE ( $\downarrow$ )	RMSE ( $\downarrow$ )	$\delta_{01}$ ( $\uparrow$ )
LH	0.1150	0.1562	97.87	0.1068	0.1459	98.40	0.1022	0.1434	98.33
TLH	0.0840	0.1221	98.98	0.0839	0.1220	98.98	0.0848	0.1227	98.96
OIH	0.0933	0.1266	99.02	0.0842	0.1150	99.40	0.0768	0.1068	99.47
R-CNN	0.0702	0.0995	99.61	0.0558	0.0809	99.78	0.0556	0.0784	99.84
R-Attn	0.0652	0.0915	99.73	0.0603	0.0851	99.78	0.0592	0.0829	99.79

Table 1: Quantitative results of different decoders based on DINOv3 backbones (ViT-S, ViT-B, ViT-L). Metrics include MAE, RMSE, and

\delta_{01}

\delta_{01}

is presented in percentage.

Figure 5: Parametric ablation study of thermal diffusion. The top and middle rows display the original RGB image and the corresponding simulated steady-state surface temperature maps under various diffusion coefficients ( $\alpha,\beta$ ).

Figure 6: Wood thermal responses across six distinct structural configurations. The top row demonstrates an idealized isotropic diffusion with coefficients $\alpha=\beta=1$ , while the bottom row incorporates spatially varying in-plane anisotropy ( $\alpha=\beta\in[0.3,3]$ ). The random $k_{z}$ distribution exhibits the lowest surface temperatures due to the thermal resistance introduced by heterogeneous layer stratification.

Figure 7: Wood internal heat transfer profiles mapping the temperature gradients from the bottom thermal boundary ( $50^{\circ}\text{C}$ ) to the top surface. The uniform and linear $k_{z}$ distributions (left and center columns) facilitate relatively continuous vertical heat conduction. In contrast, the randomly stratified distributions (right column) disrupt direct pathways, introducing tortuous thermal flow and non-linear temperature drops across the wood thickness.

Figure 8: Qualitative results of different decoders based on DINOv3 backbones (ViT-S, ViT-B, ViT-L).

Figure 9: Distributions of learned softmax layer weights across different decoders and DINOv3 backbones (ViT-S, ViT-B, ViT-L).

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Estimator	Oak sample #	Side A		Side B
Estimator	Oak sample #	$h[\text{W}\text{m}^{-2}\text{K}^{-1}]$	$T_{\infty}$ [ ${}^{\circ}\mathrm{C}$ ]	$h[\text{W}\text{m}^{-2}\text{K}^{-1}]$	$T_{\infty}$ [ ${}^{\circ}\mathrm{C}$ ]
OA	1	22.271	22.171	16.987	24.344
	2	22.660	22.151	19.558	24.945
	3	22.414	21.945	17.057	24.286
	4	25.525	22.958	18.766	23.509
ME	1	22.244	22.220	16.670	24.253

Appendix

Table 2: Estimated coefficient

h

and ambient temperature

T_{\infty}

across the four Oak samples.

Figure 10: Distributions of the system parameter,

\eta

, and the Pearson correlation coefficient,

r

, are evaluated on each sample across Poplar, Grandis-CC, and Grandis-RC datasets.

Materials Property Prediction

Materials property prediction uses data and physical models to estimate how composition, structure, and processing determine a material’s behavior.

See all 174 papers

Physics-Informed Modeling for Wood Thermal Analysis and Prediction
Jun 22, 2026Jingren Xie, Alex John Buckthal, Ryan Anthony O'Connor +2Thermal Imaging Materials Property Prediction
Technical University of Denmark Denmark · Pioneer Centre for AI Denmark · The Royal Danish Academy Denmark
Beyond Thermal Imaging: Inferring Thermophysical Properties from Time-Resolved Thermal Observations
Jul 8, 2026Chenghao Xu, Malcolm Mielle, Olga FinkThermal Imaging 3D Reconstruction
Intelligent Maintenance and Operations Systems, ´Ecole polytechnique f´ed´erale de Lausanne, Station 18, Lausanne, 1015, Vaud, Switzerland. · Schindler Lab EPFL, Schindler Holding AG, Station 18, Lausanne, 1015, Vaud, Switzerland.
Text2Thermal: Physics-Aware Thermal Image Synthesis from Textual Priors
Sep 3, 2026Tayeba Qazi, Brejesh Lall, Prerana MukherjeeThermal Imaging Conditional Image Generation
Bharti School of Telecommunications Technology and Management, Indian Institute of Technology Delhi, Delhi, India · School of Engineering, Jawaharlal Nehru University, Delhi, India

Jun 22, 2026J/K move · Enter open · S save

Physics-Informed Modeling for Wood Thermal Analysis and Prediction

Jingren Xie, Alex John Buckthal, Ryan Anthony O'Connor, Isak Worre Foged+1

Technical University of Denmark Denmark · Pioneer Centre for AI Denmark · The Royal Danish Academy Denmark

Thermal Imaging Materials Property Prediction Physics-Informed ML

Open paper AskPDF arXiv

Wood materials exhibit complex, spatially varying thermal properties that challenge traditional architectural assumptions of material homogeneity. Although data-driven approaches can directly map wood RGB images to their corresponding thermal responses, they operate as uninterpretable black boxes that prioritize statistical correlation and may absorb experimental noise rather than thermodynamic plausibility. To address these limitations, we present physics-informed deep learning frameworks that integrate partial differential equations (PDEs) to predict pixel-level thermal responses of spatially heterogeneous wood materials using wood RGB images and testbed temperature maps. Specifically, we investigate two distinct approaches to enforcing a normalized 2D steady-state heat transfer equation derived from the general heat transfer equation: Physics-Informed Convolutional Neural Networks (PICNNs), which embed physics as a soft penalty term in the loss function, and Physics-Integrated Convolutional Neural Networks (PInteCNNs), which hard-code an analytical approximator-predictor-corrector solver directly into convolutional neural networks. To validate our proposed approaches, we collect three real-world multimodal datasets of Poplar, Grandis Cross-Cut (Grandis-CC), and Grandis Radial-Cut (Grandis-RC) wood samples. We further demonstrate that embedding physical inductive biases successfully balances predictive accuracy, physical interpretability, and intra-species diversity, outperforming data-driven approaches in handling complex wood material heterogeneity and enabling the extraction of interpretable physical parameters. Project: https://zekifayes.github.io/pim

Abstract

Materials Property Prediction

Explore similar work

Physics-Informed Modeling for Wood Thermal Analysis and Prediction

Beyond Thermal Imaging: Inferring Thermophysical Properties from Time-Resolved Thermal Observations

Text2Thermal: Physics-Aware Thermal Image Synthesis from Textual Priors