Machine Learning-Driven Predictions of Crystal Symmetry Groups Using Chemical Compositions in Binary and Ternary Materials

Emphasizing the intersectionality of materials science and physics, this work investigates the profound problem of predicting crystal structures solely based on chemical compositions, a daunting task in condensed matter physics. Leveraging minimalistic, yet impactful, ionic and compositional features such as stoichiometry, ionic radii, and oxidation states, we engineered highly accurate Machine Learning (ML) classifiers capable of predicting crystallographic symmetry groups, even with the complex, multi-label, multi-class nature of the problem [1, 2]. Focusing on ternary (A_l B_m C_n) and binary (A_l B_m) materials, the developed ML models exhibit high accuracy across various symmetry groups, including crystal systems, point groups, Bravais lattices and space groups, with weighted balanced accuracies surpassing 95% even in the context of size-imbalanced data [3, 4]. This underlines that intrinsic physics is well-represented, further substantiated by illustrating the accuracy of the models aligning closely with the available data size. This pioneering approach not only provides a potential solution to an age-old problem but propels forward the expedition in discovering and developing new materials by embedding predictive analytics at its core, combining physics-informed features with data-driven methodologies.