Hierarchical input of various length - What would you do?


I would like to use 3D cad models as input in hopes to predict the cost to produce various parts(widgets). The problem is, there could be great variance in the size of the input. There is also a hierarchy to the data. The geometry is broken down in the image below (slightly simplified) .

I could use zero padding to achieve uniform input length, but wouldn’t the results suffer if the input varies to a large degree? Would a RNN be the right choice for this?

What about the hierarchy and feature tolerances? How would you maintain the relationships of the input if you need to flatten the data? I have been struggling to find information/examples for this application.