[ICLR 2025] NeFTY: Bridging Neural Fields and Differentiable Physics for 3D Thermal Tomography

TL;DR

Thermal Tomography is notoriously difficult because heat diffusion acts as a "low-pass filter," blurring internal details before they reach the surface. This paper presents NeFTY (Neural Field Thermal Tomography), a framework that solves this "un-diffusing" problem by embedding a rigorous 3D heat solver inside a neural optimization loop. Unlike previous AI methods that "guess" the physics, NeFTY enforces thermodynamic laws as hard constraints, allowing it to see through materials and reconstruct 3D defects with unprecedented clarity without needing any labeled training data.

The "Data-Fit Paradox": Why PINNs Fail at Heat

In the world of Scientific AI, Physics-Informed Neural Networks (PINNs) are the standard. However, PINNs treat physics as a "suggestion" (a soft penalty in the loss function).

The authors identify a Data-Fit Paradox: A PINN can learn to perfectly match surface temperature data while producing a completely wrong internal 3D structure. Why? Because the optimization gets stuck in "gradient stiffness"—the network prioritizes local math consistency over the global physical truth. In transient diffusion, the signal from deep defects is so faint that a standard PINN simply ignores it, settling for a "featureless" blurred solution.

Methodology: The Discretize-then-Optimize Paradigm

NeFTY abandons the soft-constraint approach of PINNs. Instead, it treats the 3D material as a Neural Field (similar to NeRF) but passes the output through a Differentiable Physics Solver.

1. Hard-Constrained Solver

Instead of asking the network to learn the temperature, NeFTY uses a Differentiable Implicit Euler solver. This ensures that at every single step of training, the generated temperature must follow Fourier’s Law of heat conduction.

2. The Adjoint Method: Making it Reachable

Simulating 3D heat over time is memory-intensive. Standard backpropagation would require terabytes of VRAM. NeFTY uses the Adjoint State Method, which allows the gradient to "flow back" through the physics without storing the entire history. This reduces memory usage from "prohibitive" to "consumer-grade GPU" levels.

Figure 1: The NeFTY pipeline. An MLP represents the diffusivity field (\alpha), which is fed into a differentiable solver to predict surface temperatures, which are then compared to real-world measurements.

3. Frequency Annealing

To capture sharp boundaries (like a crack or a void), NeFTY uses Frequency Annealing. It starts by learning the "big picture" (low-frequency thermal bulk) and gradually introduces higher-frequency details, preventing the optimization from falling into noisy local minima.

Experimental Results: Seeing the Unseen

The authors tested NeFTY on complex scenarios, including layered composite materials and samples with multiple overlapping defects.

• Accuracy: NeFTY achieved an IoU of 0.45, nearly matching supervised models that had the "luxury" of seeing the ground truth during training.
• Robustness: While 1D heuristics (like TSR) failed to account for lateral heat spread, NeFTY’s 3D physics engine naturally resolved the shape and depth of scatters.
• Generalization: Unlike U-Nets, which fail when they see a new type of defect, NeFTY works from scratch for every specimen (test-time optimization).

Figure 2: Depth-wise slices show NeFTY recovering sharp defect boundaries where standard PINNs and Grid Optimization produce noise or featureless blobs.

Critical Insights & Takeaways

The most profound takeaway is the failure of the "Sound-Only" U-Net. In NDE, we rarely have labels for every possible defect. The fact that a supervised model collapses when faced with a new defect type highlights the fragility of pure data-driven approaches in high-stakes engineering.

Current Limitations:

• Speed: NeFTY requires ~10 minutes per reconstruction because it optimizes from scratch. It is not yet "real-time."
• Contrast Range: Extremely high-contrast materials (like air vs. metal) still pose numerical challenges for gradient flow.

Conclusion

NeFTY marks a significant shift in Non-Destructive Evaluation. By merging the representational power of Neural Fields with the mathematical rigor of Differentiable Physics, it transforms thermal cameras from simple surface monitors into powerful 3D tomographic scanners.

Author Bias Note: As a technical editor, I find the "Hard Constraint" argument compelling. The "Soft Constraint" era of PINNs is hitting a wall in stiff PDE regimes, and NeFTY provides a clear, albeit computationally heavier, blueprint for the next generation of physics-AI fusion.