Added Fermi model, denoising function, dos alignment function, DOS an…

[HowWeiBin]

These are the functionalities that were requested to enhance the cookbook example. As the timeline for the cookbook is quite tight, please approve/tell me if more changes are needed ASAP

[abmazitov]

Thanks a lot @HowWeiBin ! A few comments:

1. Is there any reason to use the DOS calculation without a denoising algorithm? If it's always better and more physical to use denoising, why don't we always work with a denoised DOS and get rid of a default non-denoised setting? IMO having two settings where one is clearly told to be worse than another seems a bit confusing from a user perspective.
2. The model=True option for computing the bandgap actually falls into the same problem - why don't we always use a dedicated CNN to compute the bandgap if it's proven to be always better? What's the benefit of getting it directly from the DOS?

[ceriottm]

Thanks a lot @HowWeiBin ! A few comments:

1. Is there any reason to use the DOS calculation without a denoising algorithm? If it's always better and more physical to use denoising, why don't we always work with a denoised DOS and get rid of a default non-denoised setting? IMO having two settings where one is clearly told to be worse than another seems a bit confusing from a user perspective.
2. The model=True option for computing the bandgap actually falls into the same problem - why don't we always use a dedicated CNN to compute the bandgap if it's proven to be always better? What's the benefit of getting it directly from the DOS?

I'd be in favor of keeping the options here. Both of these are tricks to address very large errors, but make the predictions inconsistent with each other. It could very well be that when fine tuning denoising or gap head become far less preferable, and then we'd have to make a change, and we'd be discussing how we keep changing the option set of the models.

[ceriottm]

I shared some design comments offline with @HowWeiBin . I should add here that I'd call compute_DOS_and_mask_from_eigenvalues -> dos_from_eigenvalues and pad_dos_and_mask_for_training as pad_dos (and as discussed offline, make generic versions available from utils so that they can be called without having a trained model already.

[HowWeiBin]

@abmazitov and @pfebrer I have modified the code to fit the changes in the loss function in metatrain. Please review! Pol and I have removed the energy grids and we made the entire process agnostic to the energy reference

[pfebrer]

Otherwise I think that now things are much better so it looks good to me 👍

[HowWeiBin]

all tests passed! @abmazitov anything from your end?

[abmazitov]

LGTM, thank you @HowWeiBin !