We are excited (yeah, sorry) to release the code for our method and experiments on GitHub - open source matters! We’re looking forward to seeing what the community does with it, and to take another step towards universal quantum simulation with deep learning.
https://github.com/google-deepmind/ferminet/tree/main/ferminet/configs/excited
ferminet/ferminet/configs/excited at main · google-deepmind/ferminet
An implementation of the Fermionic Neural Network for ab-initio electronic structure calculations - google-deepmind/ferminet
While excited state energies are harder to calculate than ground states, double excitations - where two electrons are excited at once - are even harder. We show results on these systems that match state-of-the-art calculations at the frontier of what is possible today.
On the carbon dimer - a molecule whose excited states are responsible for the green tint of carbon-rich comets - we match experimental results *five times* more accurately than previous gold-standard calculations.
We couldn’t extend our work with neural networks to excited states because existing methods had several limitations. So we developed a new method that is general, robust, and as accurate as state-of-the-art classical methods as you scale it.
The physics of how electrons get kicked into or fall out of higher energy states, absorbing or releasing light in the process, is important for lasers, semiconductors, LEDs, solar panels, fluorescent dyes, and the biophysics of vision and photosynthesis.
We had previously shown that deep learning could be used for gold-standard calculations of the *lowest* energy states of real molecules. This is already quite challenging, but ignores all of the exotic physics that happens when molecules are stimulated.
https://www.nature.com/articles/s41570-023-00516-8
Ab initio quantum chemistry with neural-network wavefunctions - Nature Reviews Chemistry
Quantum Monte Carlo methods using neutral-network ansatzes can provide virtually exact solutions to the electronic Schrödinger equations for small systems and are comparable to conventional quantum chemistry methods when investigating systems with dozens of electrons.
I’m beyond thrilled to share that our work on using deep learning to compute excited states of molecules is out today in Science Magazine! This is the first time that deep learning has accurately solved some of the hardest problems in quantum physics.
https://www.science.org/doi/full/10.1126/science.adn0137About time to drop a paper thread. Anyone still on here?
I kind of love that almost no one on here engages with anything I post, but the one person who consistently does is a world-famous sci-fi author.
We showed that neural network VMC methods like the FermiNet can be used to accurately calculate binding energies between positrons and molecules for large challenging non-polar molecules like benzene.
The great thing about this is it was an *extremely simple* change to the base FermiNet. This shows part of why neural network VMC is so powerful - it is very easily *extensible* to all sorts of new applications.
