It isn't very easy to read DeepMind's "mind", is it? Recently the DeepMind supercomputer achieved the ability to predict protein structures to a pretty commendable level of accuracy.
To get "insight" from any physical phenomena, experiment must be able to codify a law that can be used to reproduce future experiments. But when an "Artificial intelligence" (AI) comes in the way, we get into a spot of bother because we would need to probe the "mind" of the observer, in the sense that there is an added layer of uncertainty faced by a theoretical physicist in trying to disentangle the probabilistic outcomes from the computer. This raises doubt on the reproducibility of an experiment and our understanding of a phenomenon, for example in the protein folding problem, the predicted outcome was amazingly accurate, yet we haven't really been able to reason out complex higher order interactions and energy distributions that occurs within a protein to give its characteristic structure.
Hence, are there any established methods to actually scourge out reasoning behind AI systems' functioning in physics problems that guarantees reproducibility?
 A: As far as reading DeepMind's mind, or any other AI for that matter, especially ML (machine learning) based AIs like DeepMind, it's beyond uneasy. It's impossible. It's the nature of the way they're constructed. For example, consider another ML based AI created by Google: AlphaGo. The Google engineers who made AlphaGo are barely amateur Go players themselves. Some of them might barely know how to play at all. So they don't teach it play Go; they let it loose on a surfeit of game data and let the machine sort it out for themselves. The end result is that the machine can become a brilliant player, orders of magnitude beyond the best player among the engineers who constructed it, and yet entirely useless for explaining even the simplest aspect of their own strategy.
So you're right, it's not easy to read its mind because it can't even read its own mind. It doesn't have the mysterious gift of human sentience/consciousness. (Berkeley philosopher John Searle gave a fantastic Google talk about this.)
Check out the incredible documentary on Netflix, "AlphaGo - The Movie," and you'll see: AlphaGo can beat a human to become the world champion Go player; only a human being like Lee Sodol can weep upon being unseated as world champion. AIs can't do that (yet).
