I have a decent understanding of the approach and would vote for #1. I’d say almost all applications of ML in physical sciences are #1. In contrast applying methods of statistical physics to understand how deep learning (as in DNN+SDG) works at all is a good example of #2.