Machine Learning for Alloy Composition and Process Optimization

The drive for greater efficiency in turbomachinery has led to increasingly stringent specifications for the materials used. Current methods for optimizing alloy composition and processing to meet these requirements typically rely on a combination of expert judgment and trial and error. Machine learning offers an alternative approach that leverages data resources to significantly accelerate the optimization timeline through systematic data-informed decision making. In this paper, we demonstrate the effectiveness of machine learning methods for three different alloys classes: aluminum alloys, nickel-based superalloys, and shape memory alloys. In the first two alloy classes, models are built for the alloy mechanical properties based on the composition and processing information. In the case of shape memory alloys, a model is trained to predict the austenite to marten-site transformation temperatures. In addition to achieving high baseline performance, we leverage recent methodological developments to provide well-calibrated, heteroscedastic uncertainty estimates with each prediction. By wrapping these models in an inverse design routine that takes full advantage of uncertainty information, we are able to demonstrate the feasibility of designing new alloys to meet prescribed specifications. The results indicate that this approach has the potential to fundamentally change how new structural and functional alloys are developed.