Two minute paper has a 5 minute video on a new paper:
https://www.youtube.com/watch?v=atcKO15YVD8
So, train a nural network to learn a faster representation of physics. Then just use that instead. I assume there will be a bunch of github projects on this.
"Subspace neural physics: fast data-driven interactive simulation":
https://dl.acm.org/citation.cfm?id=3340245
Non-paywalled paper available at researchgate, or this blog:
http://theorangeduck.com/page/subspace- ... simulation