Neural Field Thermal Tomography: A Differentiable Physics Framework for Non-Destructive Evaluation

TL;DR

NeFTY parameterizes the 3D diffusivity field as a continuous neural field optimized through a rigorous numerical solver, enforces thermodynamic laws as hard constraints while maintaining the memory efficiency required for high-resolution 3D tomography.

Abstract

We propose Neural Field Thermal Tomography (NeFTY), a differentiable physics framework for the quantitative 3D reconstruction of material properties from transient surface temperature measurements. While traditional thermography relies on pixel-wise 1D approximations that neglect lateral diffusion, and soft-constrained Physics-Informed Neural Networks (PINNs) often fail in transient diffusion scenarios due to gradient stiffness, NeFTY parameterizes the 3D diffusivity field as a continuous neural field optimized through a rigorous numerical solver. By leveraging a differentiable physics solver, our approach enforces thermodynamic laws as hard constraints while maintaining the memory efficiency required for high-resolution 3D tomography. Our discretize-then-optimize paradigm effectively mitigates the spectral bias and ill-posedness inherent in inverse heat conduction, enabling the recovery of subsurface defects at arbitrary scales. Experimental validation on synthetic data demonstrates that NeFTY significantly improves the accuracy of subsurface defect localization over baselines. Additional details at https://cab-lab-princeton.github.io/nefty/

Figures (15)

• Figure 1: Overview of setup. A high-speed camera enables measurements of time-resolved transient surface temperature variations following localized heating with a pulsed laser. NeFTY uses these transient measurements to reconstruct the 3D subsurface diffusivity field and reveal hidden defects.
• Figure 2: The Ill-Posedness of IHCP. Distinct internal structures (left), homogeneous vs. defective, produce nearly indistinguishable surface temperature profiles (right), illustrating the severe loss of high-frequency spatial information caused by diffusive smoothing.
• Figure 3: Overview of NeFTY. Our method combines an implicit neural representation for the 3D diffusivity field with a differentiable physics solver. The network learns the internal structure by minimizing the error between simulated and measured surface temperatures, using the adjoint method for efficient gradient backpropagation through the transient thermal simulation.
• Figure 4: Depth-wise slices of the recovered diffusivity field (Homogeneous Setting). NeFTY (Ours) successfully localizes and sizes the subsurface defects (red/blue shapes) with sharp boundaries. The PINN baseline saturates to a trivial solution due to gradient pathology. The Grid Opt. baseline is physically consistent but noisy. The Sound-Only U-Net fails to detect the OOD defects.
• Figure 5: Qualitative Results (Layered Composite Setting). Comparison of reconstruction quality in a multi-layered material. NeFTY correctly resolves both the layer transitions and the embedded defects. The baselines struggle with the complex heterogeneity. Grid Opt. introduces significant artifacts at layer interfaces, while the PINN again fails to converge to a meaningful structure.