<p>The machine learning-assisted design of new alloy compositions often relies on the physical and chemical properties of elements to describe the materials. In the present study, we propose a strategy based on an evolutionary algorithm to generate new elemental numerical descriptions for high-entropy alloys (HEAs). These newly defined descriptions significantly enhance classification accuracy, increasing it from 77% to ~97% for recognizing FCC, BCC, and dual phases, compared to traditional empirical features. Our experimental validation demonstrates that our classification model, utilizing these new elemental numerical descriptions, successfully predicted the phases of 8 out of 9 randomly selected alloys, outperforming the same model based on traditional empirical features, which correctly predicted 4 out of 9. By incorporating these descriptions derived from a simple logistic regression model, the performance of various classifiers improved by at least 15%. Moreover, these new numerical descriptions for phase classification can be directly applied to regression model predictions of HEAs, reducing the error by 22% and improving the <i>R</i>² value from 0.79 to 0.88 in hardness prediction. Testing on six different materials datasets, including ceramics and functional alloys, demonstrated that the obtained numerical descriptions achieved higher prediction precision across various properties, indicating the broad applicability of our strategy.</p>