Structure–property linkages using a data science approach: Application to a non-metallic inclusion/steel composite system

Abstract Practical multiscale materials design is contingent on the availability of robust and reliable reduced-order linkages (i.e., surrogate models) between the material internal structure and its associated macroscale properties of interest. Traditional approaches for establishing such linkages have relied largely on computationally expensive numerical simulation tools (e.g., the finite element models). This work investigates the viability of establishing low (computational) cost, data-driven, surrogate models for previously established numerical multiscale material models. This new approach comprises the following main steps: (1) generating a calibration (i.e., training) dataset using an ensemble of representative microstructures and obtaining their mechanical responses using established physics-based simulation tools (e.g., finite element models), (2) establishing objective, reduced-order, measures of the microstructures (e.g., using n-point spatial correlations and Principal Component Analysis), and (3) extracting and validating sufficiently accurate, computationally low-cost, relationships between the selected microstructure measures and effective (homogenized) properties (or performance metrics) of interest using various regression methods. In this paper, the viability of the data science approach in capturing such linkages (expressed as metamodels or surrogate models) for inelastic effective properties of composite materials is demonstrated for the first time.