You could probably use machine learning for fine-tuning the parameters of whatever fast approximation comes out of that though.

(having a plausible first-principles model and then automatically tuning the knobs until the error with a ground truth is minimised is a form of machine learning, right?)