Predictive modeling of the bainite start temperature using Bayesian inference

Abstract

The prediction of the bainite start temperature (Bs) is key to modeling the bainitic phase transformations in steels. The present work employs Bayesian inference on various existing models and presents enhanced models that allow for accurate prediction of bainite start temperatures. In a first step, a meticulously curated dataset is generated and accompanied by additional experimental Bs temperatures. Several physics-based models based on energy criteria (one diffusive and some displacive models) and a linear regression model are used to predict Bs. Adaptation and enhancement of the available models are evaluated in the framework of Bayesian inference including uncertainties, and the predictive performance is compared between the models. The concentration-dependent Gibbs energies are calculated using three different thermodynamic databases, and the models are parameterized regarding the best possible prediction of Bs. The obtained parameterization for a diffusive, some displacive and a linear regression model is used to analyze the uncertainty in the model parameters and to quantify the influence of the most important steel alloying elements on Bs. Results show that there is little difference between displacive, diffusive and data-driven approaches for prediction of Bs. The direction of the influence of main steel alloying elements is consistent with literature, and a first estimation of the effect of aluminum and cobalt is obtained. It is found that aluminum increases the bainite start temperature and the energy barrier absolute, while cobalt decreases both.