These two projects look great, and are exactly the kind of projects where I can see Google making an impact. I think it is extremely likely that we will struggle to build working reactors until we have an "AutoCAD for magnetic containment" which is able to accurately simulate real world plasma physics, after which I expect that a working design will be discovered within a few years, and then constructed within a decade or two. This follows from the simple observation that the build-test cycle for real world reactors is something like a decade, and for simulated reactors something like 1 day - a 36500X decrease in iteration time is a game changer. Such a program is very non-trivial and no one today knows how to write one.

And that's exactly the kind of stuff Google is trying to accelerate. One tangential but highly relevant example is their work on predicting molecular properties [0]. They were basically able to use neural networks to speed up the simulator by 300,000x, taking something that would take hours to less than a second. As you mention, this entirely changes the way you work and iterate. Here's a video of Jeff Dean talking more about it [1].

[0] https://ai.googleblog.com/2017/04/predicting-properties-of-m...

[1] https://www.youtube.com/watch?v=rP8CGyDbxBY&t=16m43s

How can something like this even conceptually make sense? I'm sorry, but can someone explain to me how a neural network could approximate physically-deduced laws of nature (as had-coded into the classical simulator as I imagine) "without any distinguishable error"?

This is the biggest bs I have ever heard, since an molecular physics simulator that is beat by a neural network can only just be unoptimized as hell.

I think this is an example of something similar: https://www.sciencedaily.com/releases/2019/05/190517145112.h...

> Machine learning speeds modeling of experiments aimed at capturing fusion energy on Earth