The properties of atoms, molecules and solids could all be computed reliably if we were able to solve the many-electron Schrödinger equation quickly and accurately enough.
In 2020, working with researchers from DeepMind, we introduced a new approach to this longstanding challenge. Taking heart from the success of neural networks as function approximators, we represent the many-electron wave function as a neural network we call a “FermiNet”.
The variational principle allows the network to be optimized without any need for externally generated training data. FermiNet has shown unrivalled flexibility and adaptability, whilst also producing results accurate enough to match the best scalable quantum chemical approaches.In our current PRACE grant (ends 31/3/2023), we found that FermiNet can discover quantum phase transitions (Wigner crystallisation Mott transition electron pairing) without external help.
Guided by the variational principle alone, the network converges to the correct phase spontaneously. We regard this as a game changer in materials physics.
We would now like to investigate superconductivity, 3He, excitonic liquids, solid hydrogen (with quantum protons and electrons), and metal-insulator transitions in doped semiconductors. On the technical side, we will make FermiNet capable of simulating larger systems and improve its optimisation (“training”) algorithms.