Abstract
Rapidly determining structure-property correlations in materials is animportant challenge in better understanding fundamental mechanisms and greatlyassists in materials design. In microscopy, imaging data provides a directmeasurement of the local structure, while spectroscopic measurements providerelevant functional property information. Deep kernel active learningapproaches have been utilized to rapidly map local structure to functionalproperties in microscopy experiments, but are computationally expensive formulti-dimensional and correlated output spaces. Here, we present an alternativelightweight curiosity algorithm which actively samples regions with unexploredstructure-property relations, utilizing a deep-learning based surrogate modelfor error prediction. We show that the algorithm outperforms random samplingfor predicting properties from structures, and provides a convenient tool forefficient mapping of structure-property relationships in materials science.