FDTRImageEnhancer: combining physics-informed deconvolution and microstructure-aware deep learning to enhance thermal images

Abstract We present FDTRImageEnhancer , an open-source computational framework that integrates physics-informed deconvolution with microstructure-aware deep learning for inverse thermal property mapping. The approach is demonstrated on frequency domain thermoreflectance (FDTR) phase data, but its architecture is general to problems where high-resolution structural information must be reconciled with lower-resolution physical measurements. Drawing inspiration from continuum damage mechanics, where sharp cracks are represented in a smeared or nonlocal form, we approximate FDTR’s spatial averaging as a two-parameter Gaussian convolution mimicking pump and probe laser profiles. This physics abstraction is coupled with a neural network that infers region-specific thermal conductivity, guided by structural segmentation via k -means clustering to reduce the parameter space. As a methodological proof of concept, the framework was tested using synthetic FDTR data generated from finite element simulations. Across multiple runs, it consistently identified reduced thermal conductivity at grain boundaries that were visually indistinguishable in analytically inverted, lower-resolution conductivity maps. While bulk conductivity values were recovered with high accuracy ( 0.5 % error), the model overestimated the grain boundary conductivity reduction—likely due to the simplified treatment of FDTR physics in the Gaussian abstraction. Through computational optimization, the total runtime was reduced to just a few minutes, demonstrating its potential utility in practical cases. Overall, this proof of concept establishes a foundation for future convolutional neural network-based models trained in a supervised manner on diverse FDTR datasets, capable of reproducing full FDTR physics almost instantaneously. The full Python implementation, including example datasets, is provided to enable complete reproducibility and adaptation to other inverse thermal modeling tasks.