search

choderalab / espaloma

Extensible Surrogate Potential of Ab initio Learned and Optimized by Message-passing Algorithm 🍹https://arxiv.org/abs/2010.01196
https://docs.espaloma.org/en/latest/
MIT License
212 stars 23 forks source link

Finetuning Espaloma guide? #199

Open jarvist opened 10 months ago

jarvist commented 10 months ago

Is there a guide, or any additional information, on fine-tuning Espaloma? (i.e. taking the general model, and then supplementing with some molecule-specific electronic structure calculations, retraining with early stopping or regularisation)

Similarly, is there any documentation on how the main release models are trained? I note in the release notes A manuscript describing the complete fitting process and assessment is forthcoming - but is there anything to go on while that is in preparation? I very quickly get lost amongst the QCArchive / QCFractal etc.

ijpulidos commented 9 months ago

@jarvist Thanks for your interest in our tools. You are correct, we should improve the documentation and discoverability of how we are doing these things. We appreciate you bringing this to our attention.

While not perfect, there's a good amount of documentation on how to train and evaluate espaloma from https://github.com/choderalab/refit-espaloma (this is for the latest espaloma 0.3.1 and 0.3.2 models), I hope you find this helpful.

jarvist commented 9 months ago

Thanks! Yes, that looks very useful. I think the QCArchive might be a bit of a hurdle to overcome: I imagine it would be a lot more transparent to see the actual DFT calculations, similarly this would be necessary to do any active learning (& validation at the same step).

JonathanHungerland commented 2 months ago

@jarvist I was looking for the very same thing and started to go through the guide. Did you succeed in fine-tuning towards a molecule of yours?

jarvist commented 2 months ago

No, not yet - though we've been slightly distracted by the ease of fine tuning MACE for more materials science orientated applications.

On Mon, 8 Jul 2024 at 13:10, Jonathan Hungerland @.***> wrote:

@jarvist I was looking for the very same thing and started to go through the guide. Did you succeed in fine-tuning towards a molecule of yours?

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you were mentioned.Message ID: @.***>