Domain knowledge and interpretable machine learning for designing high performance high-entropy alloys/graphene composite

The incorporation of graphene (Gr) into high-entropy alloys (HEAs) can improve the strength-plasticity trade-off relationship. However, due to the complex structure and diverse components of the materials, traditional experimental and simulation methods alone are insufficient for analyzing the effects of various factors on their mechanical properties. Machine learning (ML) techniques, which incorporate domain knowledge (DK) constraints and SHapley Additive exPlanations (SHAP) interpretability, can effectively address this challenge. In this paper, the synergistic effects of different elemental concentrations, temperature (T), Gr orientation angle (θ), and Gr volume fraction (V) on the mechanical properties of HEAs/Gr composites are examined. First, molecular dynamics (MD) simulations are employed to construct a dataset for ML, optimized through a five-step feature screening method. Subsequently, ML models for predicting ultimate tensile strength (UTS) and elongation (EL) are developed. To enhance the composite properties, parameter space reduction is achieved through the integration of DK and SHAP, combined with genetic algorithm (GA) exploration to identify optimal parameter combinations. Ultimately, two optimized composites are obtained. The optimized composites exhibit superior UTS and EL compared to the original dataset, achieving a simultaneous improvement in both strength and plasticity. Closed-loop experiments verify their accuracy through MD simulations and microstructural analysis. Our ML approach that incorporates DK and SHAP in this paper, not only validates the design strategy but also coordinates and optimizes the competing properties, providing new insights for target material design and parameter optimization.