Establishing reduced-order process-structure linkages from phase field simulations of dendritic grain growth during solidification

Establishing a rigorous quantitative relationship between solidification parameters and microstructure is critical for facilitating the fabrication of metallic components with desired properties. Conventional approaches are time and effort intensive and lack a rigorous framework for the systematic and accurate quantification of microstructure. In this study, a novel data science approach is applied to extract high-value, computationally low-cost process-structure linkages from phase field simulations of dendritic grain growth during solidification. Reduced-order measures, which are adequate for identifying the salient features of microstructures and evaluating the effects of solidification process parameters, are obtained through two-point statistics and principal component analysis. Building on these reduced-order measures, computationally efficient surrogate models are established using various regression methods. It is demonstrated that the surrogate model with the best predictive performance can provide good predictions of not only the reduced-order measures but also the two-point statistics. Moreover, it is found that some reduced-order measures exhibit very complex relations with process parameters, which results in poor prediction accuracy. Compared with discarding the features with inaccurately predicted values, including them may lead to worse recovery of higher-order microstructure statistics. These results highlight the importance of microstructure feature selection in the improvement of prediction accuracy.