Machine learning combined with solid solution strengthening model to predict hardness of high entropy alloys

Abstract

Traditional material calculation methods, such as first principles and thermodynamic simulations, have accelerated the discovery of new materials. However, it is difficult for these methods to construct models flexibly based on various target properties. And they will consume plenty of computational resources while their prediction accuracy is not good. In last decade, data-driven machine learning techniques have gradually been applied in materials science, which has accumulated a large amount of theoretical and experimental data. Machine learning is able to dig out the hidden information in these data and help to predict the properties of materials. In this work, the data source was obtained through the published references. And several performance-oriented algorithms were selected to build a prediction model for the hardness of high entropy alloys. A high entropy alloy hardness dataset containing 19 candidate features was trained, tested, and evaluated using an ensemble learning algorithm: a genetic algorithm was selected to filter the 19 candidate features to obtain an optimized feature set of 8 features; a two-stage feature selection approach was then combined with a traditional solid solution strengthening theory to optimize the features, three most representative feature parameters were chosen and then used to build a Random Forest model for hardness prediction. The prediction accuracy achieved an R2 value of 0.9416 under the ten-fold cross-validation method. To better understand the prediction mechanism, solid solution strengthening theory of the alloy was used to explain the hardness differences. Further, the atomic size, electronegativity and modulus mismatch features were found to have very important effects on the solid solution strengthening of high entropy alloys when using genetic algorithms for feature selection. The machine learning algorithm and features were also further used for prediction of solid solution strengthening properties, resulting in an R2 of 0.8811 using the ten-fold cross-validation method. These screened-out parameters have good transferability for various high entropy alloy system. In view of the poor interpretability of the random forest algorithm, the SHAP interpretable machine learning method was used to dig out the internal reasoning logic of established machine learning model and clarify the mechanism of the influence of each feature on hardness. Especially, the valence electron concentration is found to have the most significant weakening effect on the hardness of high entropy alloys.