The lack of transferable machine learning (ML) models across different material classes remains a fundamental obstacle to the predictive design of heterostructures and advanced materials. This is because conventional ML approaches, trained on single material systems, often learn system-specific, and hence nontransferable, feature–property relationships. Here, we break this paradigm by developing a feature selection strategy that explicitly prioritizes feature consensus—the agreement of descriptors across structurally distinct material families. Using lattice constant (LC) prediction as a case study, we validate this approach on cubic perovskites and spinels, which share chemical similarities but differ in geometry. By creating a unified model that accurately predicts LC for both families, we demonstrate its transferability. Crucially, through symbolic regression, we distill the consensus feature set into a simple, interpretable analytical formula that captures the underlying physics of LCs. This physically intuitive formula achieves accuracy comparable to black-box ML models, revealing that the LC is governed by a balanced interplay between ionic packing and bond coordination. Our work presents a framework that transcends system-specific models and extends to the prediction of key properties such as formation energy, demonstrating excellent cross-property transferability and opening a pathway for structure–property prediction in heterostructure and multicomponent materials design.
Gao et al. (Mon,) studied this question.