To meet rising demands in advanced applications, the pareto front of strength and ductility for steels have to be expanded. A promising steel group for going beyond the existing pareto front are carbide free bainitic steels. Our aim in this work is to use Bayesian optimization to discover novel compositions for carbide free bainitic steels and optimal heat treatment conditions to shift the known pareto front to higher values.
With Bayesian optimization, the number of necessary synthesized samples can be minimized. Bayesian optimization needs a probabilistic model as basis. Instead of a gaussian process, in this work we employ a probabilistic model as the core of the Bayesian optimization, the bainite calculator. This model connects the input parameters, composition and heat treatment, with the target properties, i.e. strength and ductility. It gives an uncertainty prediction and allows the inclusion of physical knowledge in the optimization. A modular approach allows for adaption in the case a better physical description becomes available.
To build this graphical model, knowledge in form of data and physical models are collected. Additionally, in-house synthesized samples form a trustworthy high-fidelity starting dataset. Model parameters are inferred using Markov chain Monte Carlo methods whereas gaps in the modelling framework without prior physical knowledge are completed with machine learning methods. The dataset is extended iteratively based on the predictions of the Bayesian optimization.
With this approach, we not only find a better process-structure-property-relationship of carbide free bainitic steels, but also novel carbide free bainitic steel compositions while reducing the number of needed samples substantially.