Machine-learning prediction of creep strain rate in γ/γ′ cobalt-based superalloys

Abstract

A machine-learning model was built to predict the strain rate in the steady-state regime of any γ′ strengthened Co-based superalloy as a function of temperature and stress, given inputs of alloy composition, heat treatments, and microstructure (γ′ precipitate volume fraction). The model is trained on nearly 1000 distinct γ/γ′ strengthened Co-based superalloys reported in the recent literature. We developed additional intermediary machine-learning (ML) models, requiring only a compositional input, for six materials properties: solvus-, solidus-, and liquidus temperatures, peak hardness, as well as lattice misfit and yield strength (at ambient and elevated temperature). These intermediate material properties results are fed back into the ML model to improve the accuracy of creep prediction. Finally, we validate the model by predicting intermediate properties and creep properties for 16 new alloys with an alloying element outside the training data, for which we then experimentally determine these values. The results suggest that this method produces a model which provides valuable screening data for exploring the compositional design space—even when lacking training data in that space—but it is not accurate enough to use in full replacement of experimental measurements.