Approximating QFT/Nuclear physics is a very difficult task given current algorithms, so this opens potential doors for learning "faster" approximations which allow simulating more complex systems. Of course, defining the error in this approximation becomes quite difficult. They claim the following in their paper.
1. A theoretical proof that the new operation can approximate the relevant equations.
2. An evaluation of the model trained against a traditional simulation on both equal sized and "larger sized" problems. Unfortunately they only summarize this as a "High degree of accuracy" and present charts. I'd love to see a summary table of results with a final numeric accuracy evaluation.
If you're so inclined, I think this video gives a nice description of lattice QCD:
We can use neural networks in some steps but not in others. The main use seems to be in the generation of gauge configurations (states of the Markov chains we use). We can use neural networks there because we alway do our Metropolis-Hastings accept/reject step according to the ‘true’ action, which corrects any numerical errors introduced by the ML approximation.
There are different approaches for this. I’ve been working on incorporating ML into HMC; see eg. https://arxiv.org/abs/2006.11221. For an approach based on normalizing flows, search for Shanahan eg al.
Happy to answer questions, but it’s late and I’m about to go to sleep. Feel free to ask away and I will do my best in the morning. Or email me.
"At the heart of this network structure is a novel convolutional layer that preserves gauge equivariance while forming arbitrarily shaped Wilson loops in successive bilinear layers."
Much simpler