Machine learning assisted identification of BCC-FCC regime in Al-Co-Cr-Fe-Ni

Abstract

High-Entropy Alloys (HEAs) have gathered significant attention since their first report in 2004 due to the vast compositional space offering diverse functional and mechanical properties. Quickly locating exact compositions with desired properties is essential before experimental characterization. This thesis aims to use a machine learning approach to quickly determine the compositional space of BCC, FCC, and combined BCC plus FCC phases, using the Al-Co-Cr-Fe-Ni system as a case study. A total of 75,000 compositions were generated through Latin Hypercube Sampling (LHS) and analyzed via TC-Python, resulting in 54,056 cleaned data points. Key physicochemical properties, such as mixing enthalpy (ΔHmix), Valence Electron Concentration (VEC), electronegativity difference (Δχ), and atomic size difference (δ) were found to correlate with phase formation. Machine learning models, including Decision Tree Algorithm, Logistic Regression model, Random Forest Algorithm, Support Vector Machine Classifier (SVM), K-Nearest Neighbors Classifier (KNN), and Artificial Neural Networks (ANN), were trained to predict phase formation (BCC, FCC, and BCC+FCC). Among these, the ANN model achieved the highest F1 score of 98.17%, establishing it as the best-performing model, followed by KNN, SVM, and Random Forest. Validation against literature-reported data confirmed the ANN model's accuracy at approximately 90%. For experimental verification of the identified compositional space through the ANN model, an alloy with the specified composition of Al15Co5Cr5Fe45Ni30 was cast. Results from Differential Scanning Calorimetry (DSC) and X-ray diffraction (XRD) corroborated the identified compositional range, validating the effectiveness of the machine-learning approach.