> superior method to MCMC

Superior in what ways?

I could imagine simplicity for many problems. Roughly "here, replace your complex simulation with this magic function approximator". But for problems where you're calculating a physical quantity with a true answer, I don't see NNs being "much superior".