FermiNet: Quantum Physics and Chemistry from First Principles | Hacker News Reader

FermiNet: Quantum Physics and Chemistry from First Principles

deepmind.com

122 pointshardmaru6 comments

What an unexpected surprise! Wasn't aware of the company or the work, and was somewhat bored by reading a lot of undistinguished papers applying AI to computational chemistry. But focusing on the wave function antisymmetry using a neural net is one of those ideas that is obvious when someone tells you about it, but which I had never thought of.

DeepMind is also the company behind AlphaGo, and it's now owned by Google.

AlphaFold should be mentioned too: https://deepmind.com/blog/article/AlphaFold-Using-AI-for-sci...

Also, I wonder what happened with FB Deep Learning equation solver: https://ai.facebook.com/blog/using-neural-networks-to-solve-...

Despite trying and reading it carefully I didn't understand it. I wonder if it is me or the article, or both.

I only have had entry-level introductions into QM, but had no trouble understanding this. It that may be because I do have a background in computational dynamics, but I'm no expert in either field.

If I understood correctly, what the article is trying to explain is that the software/hardware architecture optimized for neural net processing is equally suited for many-body simulation of quantum equations. The architecture allows to broadcast the intermediate results among all individual particle simulators, which is untractable in other architectures: Monte-Carlo simulations lose accuracy and coupled cluster simulations can only solve stable lattice configurations.

Personally, I like the observation they made that the fitness constraint for their training is determined by physics: whichever solution yields the lowest total-system energy wins.

Love the visualizations, anyone know where other animations like that of electron cloud motion around molecules can be found?

If this method scales to larger crystals and quantum dots, we might start seeing ML models guide materials development soon...

Reply on news.ycombinator.com