Deep Neural Networks for Ab Initio Quantum Chemistry
David Pfau, Deepmind, 12:00 ETAbstract: In this talk, I will present work on how ideas from the machine learning community can give back to computational physics, in particular deep neural networks and approximate natural gradient descent. I will present a novel deep neural network architecture, the Fermionic Neural Network (FermiNet), which can be used as an expressive class of approximate solutions (Ansätze) to the Schrödinger equation for many-electron systems. We optimize the FermiNet by Kronecker-Factorized Approximate Curvature (KFAC), an approximation to natural gradient descent which can also be used to approximate stochastic reconfiguration. This makes it possible to scale stochastic reconfiguration to Ansätze with large numbers of parameters. We show that the FermiNet is able to achieve much higher accuracy than standard variational QMC Ansätze like the Slater-Jastrow-backflow ansatz, and can exceed the accuracy of coupled cluster methods like CCSD(T) on bond-breaking systems like the transition of bicyclobutane to butadiene. This shows that deep neural networks can be used to greatly improve the accuracy of variational QMC, to the point where it is competitive with other state-of-the-art ab initio methods.