Notes:

• Experiments with Neural ODE style methods for white-light images

• Uses torchdiffeq

Pending:

• PNODEs on larger CME dataset with correct use of dataloaders

• Modified PNODEs to handle constraints such as continuity, etc.

• Training on segmented instead of raw white light images

• Setups for pushing observation images through network. Latent space loss, paired autoencoders, etc.

• Training Continuous Conv architectures for prediction of arrival time and other quantities of interest at 1au

Miscellaneous:

• Sparse CNN: https://www.cv-foundation.org/openaccess/content_cvpr_2015/papers/Liu_Sparse_Convolutional_Neural_2015_CVPR_paper.pdf

• NeuroDiffEq - https://github.com/NeuroDiffGym/neurodiffeq - general ideas are very different, somewhat closer to PINNs?

• Some analytical CFD examples in section 3 of this study? RG Link

• torch metrics like MSE, absolute loss? some experiments: Gist

For code on ME-Maverick (mostly to remain untouched):

Uses virtual environment ptvenv From home directory, do source ptvenv/bin/activate And then launch Jupyter notebook from virtual env