I wonder if there is an application for ML towards setting the per vertex blend weights. How is this process currently conducted?

Not exactly an answer to your question, but something you might find interesting.

You should check out SPIN(https://smpl.is.tue.mpg.de) and STAR (https://star.is.tue.mpg.de). You input a smaller and more controllable set of pose and shape parameters and the ML model regresses out the skeleton and associated vertices of the body. These models were trained on a large dataset of motion captured pose sequences, and aim to set a strong prior for the human body.