Presents Neural Field Thermal Tomography (NeFTY), a differentiable physics framework parameterizing 3D diffusivity as a continuous neural field for quantitative reconstruction of material properties from transient surface temperatures.
The provided context does not contain information about "Neural Field Thermal Tomography (NeFTY)" or a differentiable physics framework that parameterizes 3D diffusivity as a continuous neural field for quantitative reconstruction of material properties. While several studies discuss physics-informed neural networks (PINNs) and related deep learning approaches for reconstructing 3D temperature fields from surface data in the context of thermal tomography and non-destructive evaluation , none specifically describe a neural field-based method termed NeFTY.
Research in this domain has focused on combining convolutional neural networks (CNNs) with physical laws, such as the heat equation, to improve robustness in reconstructing internal temperature distributions under noisy or sparse measurement conditions . For instance, one approach introduces statistical uncertainty during training alongside a physics-informed loss function to enhance performance in non-ideal environments, particularly for small phantoms (e.g., 10 cm in diameter) . Another study proposes a Helmholtz-informed neural network (HINN) that transforms the time-domain heat diffusion equation into the frequency domain, enabling more accurate prediction of internal temperature fields without requiring internal measurements .
Additionally, analytical solution-embedded neural networks have been explored to reduce dimensionality and training complexity by incorporating solutions of the heat conduction equation directly into network architecture, allowing reconstruction from sparse discrete data . These methods aim to solve ill-posed inverse problems in heat transfer by integrating physical constraints, thereby reducing reliance on large labeled datasets .
However, there is no mention in the context of representing diffusivity as a continuous neural field or using coordinate-based neural representations (i.e., neural fields) for material property reconstruction within a differentiable physics framework. Therefore, based on the available information, NeFTY as described in the query is not addressed in the provided sources.
This paper introduces Neural Field Thermal Tomography (NeFTY), a novel framework designed to address the inverse problem of reconstructing internal 3D material properties from transient surface temperature measurements. Unlike traditional approaches that rely on discretized grids or voxels, NeFTY parameterizes the thermal diffusivity field as a continuous neural field using a multi-layer perceptron (MLP). By coupling this neural representation with a differentiable physics solver for the heat equation, the method enables the direct optimization of the diffusivity field from observed data. The framework leverages automatic differentiation to propagate gradients from the simulation error back through the physics model, thereby updating the neural network weights to minimize the discrepancy between simulated and actual surface temperatures.
A key contribution of this work is the seamless integration of implicit neural representations with physical constraints, which eliminates the need for explicit meshing of the material properties and avoids the resolution limitations associated with voxel-based methods. This approach allows for memory-efficient, high-resolution reconstruction of complex 3D structures and arbitrary geometries. The authors demonstrate that NeFTY can accurately recover varying diffusivity maps and detect internal defects, outperforming conventional optimization techniques that often struggle with local minima or the heavy computational costs associated with iterative finite element model updating.
This research matters significantly for the field of Non-Destructive Evaluation (NDE) as it provides a robust, quantitative tool for inspecting critical components in aerospace, civil engineering, and manufacturing. By enabling precise volumetric reconstruction from sparse surface data, NeFTY enhances the ability to detect subsurface flaws and characterize material degradation without physical intrusion. Furthermore, it exemplifies the broader potential of differentiable physics frameworks in solving complex inverse problems, offering a scalable pathway for incorporating deep learning into physics-based simulations where data fidelity and physical consistency are paramount.
This work introduces Neural Field Thermal Tomography (NeFTY), a novel differentiable physics framework for reconstructing 3D thermal diffusivity fields from transient surface temperature measurements. Unlike traditional inverse methods, NeFTY parameterizes the unknown material properties as a continuous neural field, enabling end-to-end differentiable optimization. The approach leverages the physics of heat diffusion, encoded via a neural representation, to solve the ill-posed inverse problem efficiently. By combining deep learning with classical heat transfer models, NeFTY achieves high-fidelity reconstructions without requiring explicit regularization or manual feature extraction.
The key contributions include: 1. Differentiable Physics-Informed Inversion – NeFTY integrates a neural field representation of material properties with the heat equation, allowing gradient-based optimization to recover diffusivity distributions from sparse or noisy thermal data. 2. Quantitative Non-Destructive Evaluation (NDE) – The framework enables precise material characterization without destructive testing, applicable to defect detection, composite material analysis, and thermal property mapping. 3. Generalization Across Scenarios – The method adapts to varying boundary conditions, sensor placements, and material complexities, outperforming traditional methods in reconstruction accuracy and robustness.
This work advances physics-informed neural networks for inverse problems in thermal imaging, offering a scalable and interpretable alternative to purely data-driven approaches. Its implications span industrial NDE, additive manufacturing quality control, and biomaterial characterization, where non-invasive thermal property mapping is critical. By bridging deep learning with classical heat transfer, NeFTY sets a new benchmark for differentiable inverse thermal tomography.