A Supervised Machine Learning Approach for Accelerating the Design of Particulate Composites: Application to Thermal Conductivity

In this paper, we present a supervised machine learning (ML) based computational framework for designing particulate multifunctional composite materials for desired thermal conductivity (TC). In this framework, the design variables are physical descriptors of the material microstructure to link microstructure to properties for material design. The design of experiment (DoE) based on Sobol sequence was utilized to generate a sufficiently large database for training ML models accurately. Microstructures were realized through an efficient, fast packing algorithm, and the TC of microstructures were obtained using our previous Fast Fourier Transform (FFT) homogenization method. Thereafter, the ML methods constituting a reduced-order model (ROM) was trained over the generated database to establish the complex relationship between the structure and properties. Finally, the ROM is used for material design through an optimization algorithm to solve the inverse problem of finding the material with desired properties represented by its physical descriptors. The results showed that the surrogate model is accurate in predicting the behavior of microstructure with respect to high-fidelity FFT simulations, and inverse design is robust in finding microstructure parameters according to case studies.