Machine Learning-Driven Prediction of Yield Strength in High Entropy Alloys Based on Composition

High Entropy Alloys (HEAs) are promising materials consisting of five or more elements, each contributing 5–35% of the composition, unlike traditional alloys that primarily rely on a single base element. This structure, driven by high configurational entropy, enables HEAs to maintain stable solid solutions and exhibit remarkable physical, chemical, and mechanical properties. Additionally, HEAs overcome the compositional limitations of conventional alloys, offering a broader range of compositions and combinations, with significant potential for research and applications.
Historically, materials development has depended on trial-and-error methods, which require extensive testing to achieve targeted properties, an inherently inefficient process. To address this inefficiency, some studies in metallurgy have adopted theoretical models to predict mechanical properties and inform alloy design, with examples such as the Hall-Petch relation and Vegard’s law. However, deriving universally applicable equations remains challenging due to the complexity of metallic microstructures and numerous influencing variables. Researchers often combine classical equations under specific assumptions, though such formulas frequently lack applicability across varied conditions.
This study presents an efficient, machine-learning-based methodology for predicting HEAs mechanical properties, overcoming traditional limitations. The dataset, adapted and slightly modified from Borg et al. (2020), includes 181 unique HEAs data entries, with features covering grain size, processing method, crystal structure, mechanical testing method, and elemental composition. Based on studies indicating that tree-based models outperform deep learning on small datasets, Random Forest, XGBoost, and Gradient Boosting models were employed, with Gradient Boosting demonstrating the best performance, yielding R² = 0.85 and RMSE = 192.99 MPa on the test dataset. To initially validate this model, predictions were tested on Al_xHfNbTaTiZr (x = 0, 0.3, 0.5, 0.75, 1), a composition included within the total dataset used for model development. The results demonstrated a close match with experimentally reported yield strengths, achieving a MAPE of 2.44%, which confirms the model's predictive accuracy.
To further validate the model, an additional evaluation was conducted using data not included in the total dataset. Yield strength predictions from the machine learning model were compared with values extracted from four randomly selected publications, which included the alloys CoCrFeMnNiV_x (x = 0, 0.25, 0.5, 0.75, 1), TiZrNbMo_xV (x = 0–2), Al_yCr_xFe_2-xNi_3-y (x = 0.4, 0.6, 0.8, 1; y adjusted to maintain valence electron concentration), and Al_xCoCrCuFeNi (x = 0.6, 0.7, 0.75, 0.8, 1). The results indicate that, while the model's predicted yield strength values slightly differ from reported values in the literature, it effectively captures yield strength trends according to elemental composition. The observed discrepancies may be attributed to (1) variations in processing environments, (2) differing processing techniques, (3) minor material inconsistencies, and (4) potential bias from limited training data. Nevertheless, the model’s ability to accurately predict yield strength trends suggests it could aid in identifying optimal compositions in engineering applications.
Further validation results confirm that this model is applicable across various HEA types, including Cantor alloys, RHEAs, EHEAs, and other HEAs, demonstrating its adaptability to diverse conditions. With an expanded dataset, the model could potentially offer reliable absolute yield strength predictions in addition to trend identification. Expanding this approach to predict properties like ductility could also streamline the discovery of alloys with tailored mechanical properties, providing valuable insights for advancing novel material design.