Pixel-Level Thermophysical Field Prediction from Microstructural Images Using U-Net Architectures

Composite materials, critical for thermal management in advanced engineering systems, require precise prediction of heat conduction behavior influenced by intricate microstructures. Traditional numerical methods face computational bottlenecks, while data-driven deep learning approaches offer potential for efficient microstructure-to-property mapping. This paper proposes a U-Net based deep learning framework to predict steady-state heat conduction behavior in composite materials. Leveraging a dataset comprising digital microstructural images and corresponding temperature and heat flux distributions (in $x$ and $y$ directions), the U-Net is trained to establish microstructure-thermophysical fields relationships. The U-Net achieves high accuracy, with relative errors of approximately 1 % for temperature and heat flux predictions across three output channels. Notably, the effective thermal conductivity is predicted with an error of $\text{0. 1 2 \%}$. The U-Net's encoder-decoder architecture, enhanced by skip connections and hierarchical feature learning, enables robust retention of spatial details and modeling of nonlinear thermal interactions inherent to complex microstructures. Results underscore the model's capability to bridge data-driven approaches with materials science applications, offering a computationally efficient alternative to traditional numerical methods for composites analysis. This work presents a novel approach to directly interpreting and extracting thermal physical properties from material microstructural images, streamlining traditional multi-step calculations and reducing reliance on expert interpretation.