I need to generate a 3D plane given a set of feature inputs. Most inputs are a range of values between 0 and 1 (sigmoidial), except a few. For example a rectangle:

width = 10 """in meters"""
height = 10 """meters as above"""
thickness = 1.3 """meters as above"""
chamfers = [0.2, 0.6, 0.3, 0, 0.12, 0.54, 0.2, 0.8] """edge chamfers left to right and top to bottom ordered according to origin coords, presumably 0,0"""
radiuses = [0.3, 0.7, 0.8, 0.1, 0.9, 0.23, 0.87, 0] """same as above, but for corners"""
materials = ["iron", "lead", "carbon"] """what this plane should be made of, both materials and their comb ratios are open to curiosity"""
rigidity = 0.45 """0 least rigid, 1 max rigidity"""
tensile = 0.2 """0 least strength, 1 max strength"""
"""also some interesting stuff besides pure material eng"""
reflection = 0.68 """how reflective surface is"""
heights = [[3,2,-0.54], [4,2,-0.21]] """optional map of x,y coords respective to width and height, to adjust the surface bumps"""

So it should be clear that the new plane generation using a predefined random seed data is not optimal. However it seems very difficult to determine how exactly a ML network can address this challenge.

  • What kinds of networks are more suitable for these inputs?
    • I've looked at GANs, but they seem to work best for "pixels" and "image patterns". CNNs also seem to address images, while both need to be trained using an existing dataset which tunes the outputs to the training dataset.
  • What kinds of trainings should I be looking at?

I understand you may have a lot of questions, I'll add edits addressing them. All help is appreciated, I'm beyond my skills here.

Edit 1 There will be many expected outputs, but to start off answering above questions we'd expect render data as follows:

vertices = [[1,2,3], [5,2,2], ...] """vector3 vertices for rendering"""
composition_attr = {"conductivity": <sigmoid>, "weight": <kg/m3>, ...} """for analysis, recording and passing on to other systems"""
  • $\begingroup$ What is the expected output format, i.e. the expected output datatype, so to say? Or how is the plane supposed to look like? Is it just supposed to be a flat surface which could be specified to live within four specific corner-points? $\endgroup$ – Daniel B. Jul 16 at 20:57

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.