Machine learning predictions of superalloy microstructure

TL;DR

This paper presents a machine learning approach using Gaussian process regression with a physically-informed kernel to accurately predict phase compositions in nickel-base superalloys, outperforming traditional CALPHAD methods and providing uncertainty quantification.

Contribution

The study introduces a novel Gaussian process regression model with a physically-informed kernel for superalloy microstructure prediction, offering improved accuracy and uncertainty quantification over existing methods.

Findings

Model achieves $R^2>0.8$ for most components.

Predicts $ ext{γ'}$ phase with RMSE=0.006 for benchmark alloys.

Outperforms CALPHAD in accuracy for phase composition predictions.

Abstract

Gaussian process regression machine learning with a physically-informed kernel is used to model the phase compositions of nickel-base superalloys. The model delivers good predictions for laboratory and commercial superalloys, with $R^2>0.8$ for all but two components of each of the $\gamma$ and $\gamma'$ phases, and $R^2=0.924$ ($\mathrm{RMSE}=0.063$) for the $\gamma'$ fraction. For four benchmark SX-series alloys the methodology predicts the $\gamma'$ phase composition with $\mathrm{RMSE}=0.006$ and the fraction with $\mathrm{RMSE}=0.020$, superior to the $0.007$ and $0.021$ respectively from CALPHAD. Furthermore, unlike CALPHAD Gaussian process regression quantifies the uncertainty in predictions, and can be retrained as new data becomes available.