Orbital-free density functional theory (OF-DFT) holds the promise to compute ground state molecular properties at minimal cost. Despite this, decades of research did not yield a kinetic energy functional with sufficient accuracy for general use in computational chemistry.[1]
Previously several ML models have been published that are able to predict the kinetic energy density and kinetic potential of Kohn-Sham ground state densities.[2,3] However no ML kinetic energy functional with sufficient generality for meaningful molecular calculation has been reported so far. A major computational obstacle is that it is insufficient to precisely predict the relevant quantities for the ground state only; instead it is necessary to estimate them accurately for all intermediate densities during a calculation.[3] Unfortunately, though, the kinetic potential can only be calculated from KS-DFT solutions.
As a consequence of this inverse DFT can be leveraged to calculate the external potential of those intermediate densities. This allows the calculation of the kinetic potential for those densities. Furthermore, we propose to overcome this limitation by not just training on standard KS-DFT ground states but also on ground states from perturbed external potentials. The novel equivariant network KineticNet has been trained accordingly and achieves OF-DFT calculations for two electron systems.
T. Kaczun