Fine-Tuning CHGNet with MD DFT Data and Structure-Level Properties

[CibranLopez]

Hi!

I’m exploring CHGNet's fine-tuning / retraining capabilities, and I had two questions I would like to share:

1. Incorporating Temperature Effects in MD DFT Data:

Given a dataset of Molecular Dynamics (MD) simulations, whose DFT calculations include temperature effects, is there a recommended approach for fine-tuning CHGNet with this kind of data? Although CHGNet allows for finite-temperature MD simulations, I was not able to input temperature when fine-tuning the model on the MD structures.

2. Handling Charged Systems (or Structure-Level Properties):

Given a dataset of zero-temperatue DFT geometry relaxations with charged structures (for example, point defects such as vacancies at different charge states), is it posilble to retrain CHGNet and include this charge state data? Perhaps through graph-level embeddings or by modifying the oxidation states in the input structures?

I was wondering if whether CHGNet currently supports these features directly, or if there are recommended strategies for adapting the model to handle that (eg, by modifying the AtomEmbedding class). Any advice on model modifications, data preparation, or specific code examples would be greatly appreciated!

Thank you in advance for your help.
Best regards.