Learn NURBS from Image Data

[kalebphipps]

Is your feature request related to a problem? Please describe.
ARTIST only learns NURBS from the measured heliostat data from SRAL; however, ideally, we want to reconstruct surface data from an image (or multiple images) on the receiver.

Describe the solution you'd like
A new NURBS training process should be implemented to learn the NURBS surface of a heliostat based on images on the receiver.

Technical details
Ideally, this should be implemented to allow parallel computing:

• The images should be in a torch.Dataset, with the considered incident_ray_direction as the input and the resulting image as a torch.Tensor as the expected output label.
• Once the surface has been learnt, the NURBS parameters should be extracted in a list of FacetConfig objects so that these can be directly saved into a scenario HDF5 file.

Additional context
In the long term we also need to consider how this can be efficiently performed for multiple heliostats.

[kalebphipps]

This Issue is now part of the ARTIST Release 1.0 milestone and should be implemented by November 7, 2024 @Filos1992

Additionally, the following issues should be solved while implementing this feature:

• The default epsilon value for NURBS should be adjusted; see Issue Adjust Default Epsilon Value for NURBS #65.
• The optimizer and scheduler values should be included as parameters and not hard-coded in NURBS; see Issue Optimizer and scheduler parameters NURBS #63.