Finetuning

[DavideTisi]

my todo (sofiia):

• clean up
• finish example of simple finetuning
• add some plots or chemiscope
• add example with nc force learning(pretraining on nc, finetuing on c)
• verify if recipe generation time can be reduced

[github-actions]

Here is the build version of the cookbook in this pull request: documentation.zip, you can view it locally by unzipping documentation.zip and open the index.html with your favorite browser.

[ceriottm]

Very nice. A few comments for improvement:

1. would be nice to show training curve (train and validation error vs epoch). also it's not clear if ethanol is computed with a different flavor of DFT, you should indicate that.
2. It'd be nice (even if it make the model slower) demonstrate the accuracy of a model trained "from scratch" for 1000 epochs. the comparison between training curves would make a nice fig. 1
3. I'd also show training curves for the conservative fine-tuning, maybe showing the "appended" curve plot as in the paper
4. another nice thing would be to show in the final parity plot the last point in the NC pre-training in the background with low opacity, so it's clear how much it improved in just 50 epochs.