Physics-Guided Machine Learning for Predicting Lattice Parameters of Crystals

Machine learning (ML), as applied in many computer-aided scientific domains, often relies on ''black-box'' approaches, which sacrifice interpretability and may lead to computational artifacts and erroneous conclusions. Addressing this issue, physics-guided machine learning (PGML) successfully demonstrated its effectiveness in crystallographic group prediction [1, 2, 3]. Building upon this foundation, we present a PGML framework that predicts lattice parameters of binary and ternary compounds directly from their chemical compositions. The model integrates domain knowledge by using physically meaningful elemental descriptors, such as ionic radii and oxidation state, to ensure interpretability. Using publicly available crystallographic datasets, we develop prototype-specific models for different crystal systems. The PGML framework achieves an overall test prediction error below 5% compared to the reference values, demonstrating excellent accuracy. These results show that integrating physics into ML models enables accurate and interpretable prediction of structural properties, offering a promising route toward accelerated materials discovery.