Despite significant research efforts through programs such as the Materials Genome Initiative, the quantitative design and discovery of engineering materials remain a challenging and slow process. The primary roadblocks remain the complexity of the design spaces involved (e.g., the space of material microstructures) and the cost of traditional explorative tools (e.g., physics-based simulations). Our approach integrates a machine learning framework, leveraging several important tools from Materials Informatics, with microstructure-explicit simulations to rapidly explore these high-dimensional spaces.
Two engineering material systems are considered. The first is an Al2O3-TiB2 composite system, where we study the effects of microstructure on its failure mechanisms and energy dissipation capacity. The second is a polymer-bonded explosive system, where we investigate the influence of the microstructure on its shock-to-detonation transition behavior by quantifying the dependence of the run distance to detonation on the shock pressure.
The framework facilitates the rapid establishment of a learned structure-property relationship in a manner that enables materials design and engineering while minimizing the number of expensive simulations necessary. To achieve the objective, our approach integrates three toolsets: (1) quantification and feature engineering using two-point correlation functions for topologically compact statistical microstructure description, along with Principal Component Analysis for dimensionality reduction of correlation maps; (2) synthetic microstructure generation for rapid identification of possible candidate structures; and (3) active learning on Multi-Output Gaussian Process Regression model for rapid establishment of structure-property relations to minimize the number of necessary simulations. The framework provides a versatile toolset for multiscale materials design.