O-kRystal: an orbital learning model predicts more accurate band structures than standard DFT for nearly 6,000 materials

Using a low-cost PBE calculation, it estimates the HSE06 level energies with an average error of 0.171 eV, and reduces the error by 68.7 per cent for strongly correlated materials.

Knowing precisely the electronic tape structure of a material is essential to predict whether it will drive electricity, absorb light or magnetize itself, but the most faithful calculation methods are very expensive. A team of William Goddard and Anima Anandkumar (Caltech) presents O-kRystal, which starts from a low-cost density functional theory (PBE) and uses it both as an energy base and as a source of orbital information, both in real space and in reciprocal space.

Correct status by state

The model predicts for each electronic state a correction dependent on crystalline moment and spin. For states less than 5 eV from the middle of the prohibited band, it recovers the energies of the hybrid method HSE06 with an average absolute error of 0.171 eV over nearly 6,000 three-dimensional, insulating, metal or magnetic materials. Of six highly correlated materials, it reduces PBE error by 68.7%. Driven only on points of high symmetry, it remains precise elsewhere along band paths, and extends to energies of near-particle GW for two-dimensional materials.