Bayesian approach in predicting mechanical properties of materials: Application to dual phase steels

An essential task in materials science and engineering is in quantifying the linkages between physical variables of a material to its properties. These linkages are both complex and computationally expensive to quantify, as evidenced by rigorous modeling efforts and time-consuming simulations. Hence, practicality dictates that tasks such as materials design that require numerous evaluations are largely limited to qualitative assessment or traditional trial and error. In this work, microstructure-based simulations with model parameters calibrated to reproduce experimental data are employed to make a qualitative assessment of how physical variables of dual-phase steel are correlated to its properties. Afterward, the linkages between physical variables of dual phase steel to its property are computed with a limited amount of microstructure-based simulation data by adopting the Bayesian approach, namely Gaussian process regression (GPR). Even with a small amount of data, GPR yielded an impressive level of accuracy. Furthermore, because microstructure-based simulations are based on experimental data, the quantified linkages are transferable to experimental data.