Theory Seminar: * FermiNet: learning ab initio many-electron wavefunctions *

James Spencer
The many-electron Schrödinger equation allows us, at least in principle, to understand much of chemistry from first principles.

James Spencer
Group Seminar via Zoom
Wed, 15. September 2021, 11:30 am (MEZ)

Abstract:

The many-electron Schrödinger equation allows us, at least in principle, to understand much of chemistry from first principles. Unfortunately exact solutions for almost all systems are NP-hard and so are out of reach. Substantial success has been made by pursuing polynomially-scaling approximations. The key trade-off, as ever, is between computational efficiency and accuracy, which are governed by the functional form used to approximate the wavefunction. Neural networks have shown impressive power as accurate practical function approximators and promise as a compact wave-function Ansatz for spin systems. Problems in electronic structure require wave functions that obey Fermi-Dirac statistics, where the wavefunction must be anti-symmetric with respect to exchange of any two electrons. FermiNet is a neural network architecture which directly imposes this requirement and can be trained with a standard variational Monte Carlo algorithm. Using no data except for the nuclear charges and positions, we demonstrate that a neural network can be trained to represent wavefunctions of complex chemical systems to excellent accuracy, surpassing conventionalvvariational Monte Carlo methods, and opening the possibility of accurate direct optimization of wave functions for previously intractable many-electron systems.

If you'd like to participate in the seminar, please contact us!