Pre-trained model weights for ODAC23 IS2RS.

[hspark1212]

Thank you for making the great repo for ODAC dataset.

When I tried to find pretrained models for IS2SR task in the ODAC23 at MODELS.md, I was not able to find the models for IS2SR task while there are pre-trained models for S2EF and IS2RE.

It would be appreciated to inform me how to find the model in the repo.

Best, Hyunsoo Park.

[anuroopsriram]

@hspark1212 The IS2RS task is solved by using the S2EF models and running structure relaxations using them. The S2EF models are equivalent to ML force fields, so they can be used with an optimizer to iteratively update the atomic positions to minimize energies. This in turn solves the IS2RS task.

This notebook explains how to run the relaxations: https://github.com/Open-Catalyst-Project/ocp/blob/main/tutorials/OCP_Tutorial.ipynb

I'll update the MODELS.md file to clarify this. Please reopen this task if you have any other questions.

[hspark1212]

Thank you for your prompt response!

I now understand how to execute the relaxation process using the S2EF models.

(1) When I reviewed the ODAC paper, I noticed that there was also an IS2RS task apart from S2EF and IS2RE. I'm curious to know if the IS2RS task was implemented by employing the same relaxation method using S2EF models and optimizers as outlined in the tutorial. (2) The S2EF models were originally pretrained to predict energy and forces based on input structures and gas molecules. Consequently, I'm interested in knowing whether these S2EF models are capable of accurately predicting energy and forces when only structural information is provided as input.

Best regards,

[anuroopsriram]

1. Yes, we used the same relaxation procedure for the IS2RS task.
2. For any ML model, one should carefully verify its performance when testing outside the training domain. If I had to guess, I would say that the force predictions could be good if the MOF is close to its ground state because all of our training data started with a pre-relaxed MOF. The energy predictions will probably not be as accurate because the models were specifically trained to predict adsorption energies.