It's really hard, that's why almost nobody knows how to do it :P.

Roughly speaking, the solution for me was to develop deep intuition in a lot of different domains, observe the features common to the intuitions in different domains, and abstract the common features out.

Finding the common features was very difficult, as there are a huge number of confounding factors that mask over the underlying commonalities. But it makes sense in hindsight - we wouldn't be able to develop deep intuitions in so many different domains if not for there being subtle underlying commonalities - there weren't evolutionary selective pressures specifically for the ability to develop general relativity and quantum field theory - the fact that it's possible for us means that the relevant pattern recognition abilities are closely related to the ones used in social contexts, etc.