Statistics

Machine learning

Vickers hardness test

Metallurgy

Materials science

Standard deviation

Alloy

High entropy alloys

Interpretability

Mathematics

Computer science

Microstructure

Acta Materialia

Trapped by time-consuming traditional trial-and-error methods and vast untapped composition space, efficiently discovering novel high entropy alloys (HEAs) with exceptional performance remains a great challenge. Herein, we present a machine learning-based alloy design system (MADS) to facilitate the rational design of HEAs with enhanced hardness. Initially, a hardness database was constructed, and then the key features affecting the hardness of HEAs were screened out by performing a four-step feature selection. Five descriptors including the average deviation of the atomic weight (ADAW), the average deviation of the column (ADC), the average deviation of the specific volume (ADSV), the valance electron concentration (VEC), and the mean melting point (Tm) were identified as the key features related to the hardness of as-cast HEAs. Furtherly, a hardness prediction model based on the support vector machine was constructed with the five features as the inputs. The Pearson correlation coefficients of the well-trained model reach 0.94 for both the testing set and the leave-one-out cross validation (LOOCV). Subsequently, several optimized compositions recommended by inverse projection and high-throughput screening were synthesized by experiments. The best performer exhibits ultra-high hardness, which is 24.8% higher than the highest one in the original dataset. Moreover, the Shapley additive explanation (SHAP) was introduced to boost the model interpretability, which manifests that VEC plays an important role in the prediction for hardness. Notably, VEC has a positive effect on hardness when VEC < 7.5.

A machine learning-based alloy design system to facilitate the rational design of high entropy alloys with enhanced hardness