Constructing generalizable microstructure–property maps across diverse microstructure classes

Abstract Establishing microstructure–property correlations that generalize across diverse microstructural classes remains a critical challenge in data-driven materials design. In this work, we evaluate the potential to extrapolate predictive models trained on one microstructure type (e.g., spinodal) to others (e.g., dendritic), using three distinct featurization strategies: two-point correlation functions, graph-based descriptors, and deep neural network embeddings. Our findings reveal that the Wasserstein distance is an excellent metric that correlates well with generalizability, serving as a model-agnostic yet data-aware signature of generalizability. Furthermore, we demonstrate that featurizations that conserve key microstructural features generalize better. Graphic abstract