Predicting the property contour-map and optimum composition of Cu-Co-Si alloys via machine learning

Establishing the property contour-map using the classical physics-based models is easy to fall into the dimension disaster due to the complexity of experimental variables (composition and heat treatment), limiting the discovery of alloys with high hardness and good electrical conductivity (EC). Here, we propose the combination of a Gaussian process regression-based machine learning (ML) surrogate model and the phase diagram to search for compositions with better target properties in Cu-Co-Si alloys. The ML models trained from discarded experimental data with undesirable hardness or EC establish the connection between input variables (experimental variables and other features) and target properties. After five iteration loops, the highest EC and hardness of the alloys were optimized to 69.24% IACS and 207.39 HV, 20.8% and 14.7% higher than the best values in the initial training data. The optimum compositions of Cu-Co-Si alloys are discovered from the contour maps which consist of the properties of 12,000 alloys predicted by our ML models. Combined with the phase diagram and microstructure characterization, we infer that the moderate precipitation of Co2Si leads to high conductivity and the precipitation of Co2Si and CoSi strengthens the Cu matrix.

Automated summary

This study proposes a method that combines Gaussian process regression-based machine learning and phase diagrams to search for better compositions in Cu-Co-Si alloys. By training machine learning models, the electrical conductivity and hardness of the alloys were optimized, and the optimal compositions were discovered through contour maps.