minsr
A Kaczmarz-inspired approach to accelerate the optimization of neural network wavefunctions
Goldshlager, Gil, Abrahamsen, Nilin, Lin, Lin
For many chemical properties, it suffices to work within the Born-Oppenheimer approximation, in which the nuclei are viewed as classical point charges and only the electrons exhibit quantum-mechanical behavior. The study of chemistry through this lens is known as electronic structure theory. Within electronic structure theory, methods to model the many-body electron wavefunction include Hartree-Fock theory, configuration interaction methods, and coupled cluster theory. A typical ansatz for such methods is a sum of Slater determinants which represent antisymmetrized products of single-particle states. The benefit of such an ansatz is that the energy and other properties of the wavefunction can be evaluated analytically from pre-computed few-particle integrals. Another approach to the electronic structure problem is the variational Monte Carlo method (VMC) [1, 2]. In VMC, the properties of the wavefunction are calculated using Monte Carlo sampling rather than direct numerical integration, and the energy is variationally minimized through a stochastic optimization procedure. This increases the cost of the calculations, especially when high accuracy is required, but it enables the use of much more general ansatzes.