Publisher Correction:Crystal structure prediction at finite temperatures

Correction to:npj Computational Materials https://doi.org/10.1038/s41524-023-01120-6,published online 21 October 2023 Current caption to Fig.4 reads“P–T phase diagram of MgSiO_(3)”.The correct caption should be“P–T phase diagram of Al”.The original article has been corrected.

Crystal structure prediction at finite temperatures

Crystal structure prediction is a central problem of crystallography and materials science, which until mid-2000s was consideredintractable. Several methods, based on either energy landscape exploration or, more commonly, global optimization, largely solvedthis problem and enabled fully non-empirical computational materials discovery. A major shortcoming is that, to avoid expensivecalculations of the entropy, crystal structure prediction was done at zero Kelvin, reducing to the search for the global minimum ofthe enthalpy rather than the free energy. As a consequence, high-temperature phases (especially those which are not quenchableto zero temperature) could be missed. Here we develop an accurate and affordable solution, enabling crystal structure prediction atfinite temperatures. Structure relaxation and fully anharmonic free energy calculations are done by molecular dynamics with aforcefield (which can be anything from a parametric forcefield for simpler cases to a trained on-the-fly machine learning interatomicpotential), the errors of which are corrected using thermodynamic perturbation theory to yield accurate results with full ab initioaccuracy. We illustrate this method by applications to metals (probing the P–T phase diagram of Al and Fe), a refractory covalentsolid (WB), an Earth-forming silicate MgSiO_(3) (at pressures and temperatures of the Earth’s lower mantle), and ceramic oxide HfO_(2).

Graph neural network guided design of novel deep-ultraviolet optical materials with high birefringence

Finding crystals with high birefringence(Δn),especially in deep-ultraviolet(DUV)regions,is important for developing polarization devices such as optical fiber sensors.Such materials are usually discovered using experimental techniques,which are costly and inefficient for a large-scale screening.Herein,we collected a database of crystal structures and their optical properties and trained atomistic line graph neural network to predict theirΔn.To estimate the level of confidence of the trained model on new data,D-optimality criterion was implemented.Using trained graph neural network,we searched for novel materials with highΔn in the Materials Project database and discovered two new DUV birefringent candidates:NaYCO_(3)F_(2) and SClO_(2)F,with highΔn values of 0.202 and 0.101 at 1064 nm,respectively.Further analysis reveals that strongly anisotropic units with various anions andπ-conjugated planar groups are beneficial for highΔn.