The overall thermal property of composites is significantly influenced by both the structure of the composite and its interface characteristics. Hence, the establishment of a structure-property relation (SPR) is crucial for designing novel materials with desired properties. However, challenges arise in identifying suitable descriptors that accurately represent the material microstructure and obtaining costly training data, particularly when derived from experimental measurements. In this research, we address this challenge using Machine Learning (ML) techniques on image-based microstructure and thermal conductivity data, which are predicted from physical simulations and are validated with experimental measurements. The input microstructure for the simulations is reconstructed from scanning electron microscope (SEM) image of the material. Diffuse interface method is used to interpolate the phase and interface properties, which is advantage in managing complex microstructures [1,2,3].

Additionally, a ML-based SPR is developed to predict the effective thermal conductivity of new microstructures (shown in Figure 1). To train the model, 1440 synthetic microstructural images are generated to closely resemble SEM micrograph of the material. Principal Component Analysis (PCA) is utilized to condense the 2-point statistics, serving as the statistical representation of the generated microstructure. These condensed statistics are then used as inputs for the ML model. Besides, the physical model computes the effective thermal conductivity for all synthetic microstructures, and these values are provided as outputs to the model. The trained model shows excellent performance, applicable for tasks such as sensitivity analysis and the inverse design of optimal materials [1].