The spatial arrangement of compositions in functional materials plays a pivotal role in modulating their properties. While various representations exist for predicting the benchmark properties of constituent compositions, geometric factors like heterogeneity beyond the atomic scale, which are crucial to the overall behavior of the material, are not encoded explicitly, leading to poor generalization to materials with scarce training data. In this work, we introduce a novel representation, called a space series, which explicitly encodes the geometric features of heterostructures. As a proof of concept, we apply this approach to the property prediction of core-shell quantum dots (CSQDs). Our model, designed to generate and leverage these geometric-compositional sequences, outperforms traditional machine learning methods in terms of generalization across different data sets. Further analysis reveals that such sequences enable the model to effectively capture spatial dependencies within the system, offering a novel pathway for incorporating geometric factors into functional material predictions.
Deng et al. (Sun,) studied this question.