Correlation analysis of materials properties by machine learning: Illustrated with stacking fault energy from first-principles calculations in dilute fcc-based alloys

Advances in machine learning (ML), especially in the cooperation between ML predictions, density functional theory (DFT) based first-principles calculations, and experimental verification are emerging as a key part of a new paradigm to understand fundamentals, verify, analyze, and predict data, and design and discover materials. Taking stacking fault energy (γ SFE) as an example, we perform a correlation analysis of γ SFE in dilute Al-, Ni-, and Pt-based alloys by descriptors and ML algorithms. These γ SFE values were predicted by DFT-based alias shear deformation approach, and up to 49 elemental descriptors and 21 regression algorithms were examined. The present work indicates that (i) the variation of γ SFE affected by alloying elements can be quantified through 14 elemental attributes based on their statistical significances to decrease the mean absolute error (MAE) in ML predictions, and in particular, the number of p valence electrons, a descriptor second only to the covalent radius in importance to model performance, is unexpected; (ii) the alloys with elements close to Ni and Co in the periodic table possess higher γ SFE values; (iii) the top four outliers of DFT predictions of γ SFE are for the alloys of Al23La, Pt23Au, Ni23Co, and Al23Be based on the analyses of statistical differences between DFT and ML predictions; and (iv) the best ML model to predict γ SFE is produced by Gaussian process regression with an average MAE < 8 mJ m-2. Beyond detailed analysis of the Al-, Ni-, and Pt-based alloys, we also predict the γ SFE values using the present ML models in other fcc-based dilute alloys (i.e., Cu, Ag, Au, Rh, Pd, and Ir) with the expected MAE < 17 mJ m-2 and observe similar effects of alloying elements on γ SFE as those in Pt23X or Ni23X.