FDTRImageEnhancer: Combining Physics-Informed Deconvolution and Microstructure-Aware Deep Learning to Enhance Thermal Images

TL;DR

FDTRImageEnhancer addresses the challenge of reconciling high-resolution structural information with lower-resolution FDTR phase data to map local thermal conductivities, especially across grain boundaries. The authors fuse a physics-informed Gaussian convolution surrogate with a microstructure-aware neural network, reducing parameter complexity via region clustering and basing training on precomputed physics through a surrogate loss that compares predicted and analytically inverted conductivities. Demonstrations on FEM-generated synthetic data show bulk conductivities recovered with sub-0.5% error and qualitative recovery of grain-boundary conductivity drops obscured in conventional inversions, highlighting the method’s potential to reveal microstructure–thermal transport interactions. The work provides an open-source, adaptable framework for inverse thermal transport problems, with clear paths for improving convolution fidelity and extending to other material systems and imaging modalities.

Abstract

We present FDTRImageEnhancer, an open-source computational framework that improves thermal conductivity mapping from Frequency Domain ThermoReflectance (FDTR) phase data by integrating a physics-based Gaussian convolution abstraction with microstructure-aware deep learning. The Gaussian kernel models the spatial averaging effects of pump and probe beams, while k-means clustering of high-resolution structural images reduces the parameter space for inverse modeling. A physics-informed neural network jointly minimizes phase-data error and deviation from analytically recovered conductivity maps, enabling the detection of grain boundary thermal conductivity drops visually obscured in conventional FDTR inversions. Demonstrated on finite element-generated synthetic data, the framework recovers bulk values within less than 0.5% error and qualitatively resolves grain boundary effects despite limited image resolution. Full Python code and datasets are provided for reproducibility, with the methodology readily adaptable to other inverse thermal transport problems.

Figures (20)

• Figure 1: 1D Bar initially heated at the middle
• Figure 2: Neural network schematic
• Figure 3: Temperature profile evolution comparison
• Figure 4: FDTR Experiment Setup
• Figure 5: FEM "Ground Truth"