Search strategy for rare microstructure to optimize material properties of filled rubber using machine learning based simulation

T Kojima and T Washio and S Hara and M Koishi, COMPUTATIONAL MATERIALS SCIENCE, 204, 111207 (2022).

DOI: 10.1016/j.commatsci.2022.111207

A shortcut to understand microstructure-property relationship, such as relationship between filler morphology and its modulus, is sampling and analysis of microstructures that induce the desired material property. However, the morphologies that induces the desired property, e.g. extremely high modulus, are very limited and hard to be searched by simple random sampling. Particularly, in the case of filled rubber, the simulation of complex filler morphology involves hundreds of filler particles. This makes the random sampling infeasible, because the number of parameters is O(3n) when using coordinates of the n particles as the search objective. Recent advancement of the sampling methods reported as a part of the materials informatics remains efficient sampling of the targeted microstructure characterized by 10 parameters at most. In this paper, we propose a novel and effective three-step search method to efficiently sample the filler morphology inducing extremely high modulus in a several hundred dimensional parameter space. We demonstrated the efficiency of the proposed method through the comparison with the random sampling in 750 dimensional parameter space for obtaining the morphologies providing the extremely high modulus. Half of the sampled morphologies provided higher stresses than the top 0.1% morphologies found by the random sampling. This is the first work that developed the efficient sampling method of complex material microstructures having targeted properties in a very high dimensional parameter space.

Return to Publications page