We need a new modelling pipeline for the ice rheology inversions. Since the inversions will be based on the average ice velocities for 2017-2018 from Millan et al. (2022), we don't need a solver, and therefore we don't need to use a UDE.

Most of the work is already there, but some new features are needed:

• Think of which features will be relevant to infer A and C from the SIA. Download and add that data as input features of the NN.
• Re-design the NN architecture and hyperparameters with something more suitable for this task. The one we're using for the UDE is probably way too simple.
• Decide on the spatialization of the inversions. C will most likely have to be inverted in a distributed matter, i.e., an individual C for each pixel. Whereas A will probably make more sense as a unique value per glacier.
• Anticipate issues related to equifinality. Try to constrain A in a more aggressive way, penalizing the use of A for fast surface velocities, which should in theory be described by C.