Exploring the design space of discontinuous metal matrix composites through domain-knowledge enhanced machine learning

Abstract

Tailored reinforcement architectures in discontinuous metal matrix composites (DMMCs) offer superior mechanical performance with broad scientific and financial interests. This study presents a domain-knowledge enhanced machine learning approach to efficiently explore the design space of Al-SiC DMMCs for optimization. A substantial dataset containing 140,000 instances, resembling characteristic reinforcement configurations and variants, is generated using a series of algorithms. Employing high-throughput finite element analysis, the elastic properties of each configuration are estimated. Statistical analysis reveals that a more homogeneous distributed reinforcement contributes to mechanical stability, whereas configurations with extreme performance tend to have inhomogeneous reinforcement distribution. A deep residual neural network trained on this dataset accurately learns the structure-property correlations. Coupled with a genetic algorithm, the framework identifies optimal configurations across different volume fractions for maximizing/minimizing properties including tensile modulus, shear modulus, and Poisson's ratio. Comparative analysis shows the incorporation of domain knowledge improves data quality, facilitating more effective design space exploration. This study contributes to advancing composite materials design, particularly for next-generation high-performance DMMCs.