Property Prediction Troubleshooting

[Dreyes13-work]

I have recreated conditioned property predictions (bandgap, bulk modulus, etc.) for x number of properties. After it goes through MatterSim we get the energy, force, and stress of these materials. My question is: Is there somewhere within the python blocks that we can append the code to see the properties and view them as statistical data after predictions are made or must we use something like CGNN to view that data? I figured since we condition on specific properties, we would be able to see those properties directly after predictions are done. Does this make sense, or did I miss something in the repository?

[luisbro]

Hey, when you say you recreated conditioned property predictions, do you mean you finetuned the model for conditioned generation, like described in the readme? Because as far as I know, MatterGen only allows for generation, but not prediction of any properties and MatterGen is trained completely independently of MatterSim. If you would like to have a prediction for how large the band gap is, you would have to train MatterSim to do this, which is currently not provided in the public code of MatterSim (also mentioned in microsoft/mattersim#115, microsoft/mattersim#108)

[Dreyes13-work]

That is exactly what I mean, I am referring to conditioned generation. Also, yes MatterGen only allows for generation, while MatterSim is the evaluation or relaxation of those generated materials. MatterSim provided specific properties that cannot be physically duplicated in a lab alone (stress, force, and energy). Hence, why I need to figure out how to extract or simply obtain those properties that provide the full spectrum of each generated material. I attempted CGCNN but fell short with a 404 code with a missing pretrained dataset that no longer is available in the repo. What are your thoughts on the angle I should take next? Appending the MatterSim code? A different NN predictor? Thank you for your response!

…

-Danny Reyes

________________________________ From: Luis ***@***.***> Sent: Sunday, May 25, 2025 9:24 AM To: microsoft/mattergen ***@***.***> Cc: Danny Reyes ***@***.***>; Author ***@***.***> Subject: Re: [microsoft/mattergen] Property Prediction Troubleshooting (Discussion #160) You don't often get email from ***@***.*** Learn why this is important<https://aka.ms/LearnAboutSenderIdentification> WARNING This email originated external to the NMSU email system. Do not click on links or open attachments unless you are sure the content is safe. Hey, when you say you recreated conditioned property predictions, do you mean you finetuned the model for conditioned generation, like described in the readme? Because as far as I know, MatterGen only allows for generation, but not prediction of any properties and MatterGen is trained completely independently of MatterSim. If you would like to have a prediction for how large the band gap is, you would have to train MatterSim to do this, which is currently not provided in the public code of MatterSim (also mentioned in microsoft/mattersim#115<microsoft/mattersim#115>, microsoft/mattersim#108<microsoft/mattersim#108>) — Reply to this email directly, view it on GitHub<#160 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/BJK65LFHTWTVQIEASIIPLHT3AHOCXAVCNFSM6AAAAAB5UJAF4GVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTGMRWGQZTIMI>. You are receiving this because you authored the thread.Message ID: ***@***.***>

[luisbro]

I might not have understood your question fully. Are you interested in predicting a property like the band gap for generated materials or are you interested in the predicted values from MatterSim, including forces, stress and energy?

If you want MatterSim to predict forces stress and energy then there is the summarized output from the evaluate command in metrics.json

mattergen-evaluate --structures_path=$RESULTS_PATH --relax=True --structure_matcher='disordered' --save_as="$RESULTS_PATH/metrics.json"

If you want to see it for each structure individually, you can look in here

mattergen/mattergen/evaluation/evaluate.py

Lines 61 to 65 in ec029d1

	return evaluator.compute_metrics(
	metrics=evaluator.available_metrics,
	save_as=save_as,
	pretty_print=True,
	)

and exchange the compute_metrics with another function I saw, as_dataframe, which doesnt compute a summary, but lists everything individually.

If you would like to predict a property like the band gap, which isn't natively supported by MatterSim, I can't really give good advice on that :/ you probably have to train some part of it on your own

[Dreyes13-work]

I apologize for not being fully transparent. I am interested in predicting a property like the band gap for generated materials. I have successfully run MatterGen and MatterSim and am able to view many attributes to all my generated materials such as symmetry, crystal coordinates, stress, force, energy, chemical comp, etc.. My issue is in my line of work we want properties such as bulk modulus, bandgap, etc., or these predicted materials don't hold much value to us. Conditioning on such properties is understood but what those values are post MatterGen and post MatterSim for those properties is what I am seeking to find out. Yes, I am going to attempt my own training next. As it's required to have some pretrained set of data to evaluate the newly generated material properties through some neural network like CGCNN. I only mention CGCNN as some of its core contributors are the same as those on MatterGen and MatterSim. But they have limited the public to what we have access to, as the repo does not provide us 100% of the workflow. This conversation is very helpful and again I thank you.

…

-Danny Reyes Sent from Samsung Galaxy smartphone. Get Outlook for Android<https://aka.ms/AAb9ysg>

________________________________ From: Luis ***@***.***> Sent: Sunday, May 25, 2025 11:23:02 AM To: microsoft/mattergen ***@***.***> Cc: Danny Reyes ***@***.***>; Author ***@***.***> Subject: Re: [microsoft/mattergen] Property Prediction Troubleshooting (Discussion #160) You don't often get email from ***@***.*** Learn why this is important<https://aka.ms/LearnAboutSenderIdentification> WARNING This email originated external to the NMSU email system. Do not click on links or open attachments unless you are sure the content is safe. I might not have understood your question fully. Are you interested in predicting a property like the band gap for generated materials or are you interested in the predicted values from MatterSim, including forces, stress and energy? If you want MatterSim to predict forces stress and energy then there is the summarized output from the evaluate command in metrics.json mattergen-evaluate --structures_path=$RESULTS_PATH --relax=True --structure_matcher='disordered' --save_as="$RESULTS_PATH/metrics.json" If you want to see it for each structure individually, you can look in here https://github.com/microsoft/mattergen/blob/ec029d177c93709fa9a2ea4e48b872760d09c63b/mattergen/evaluation/evaluate.py#L61-L65 and exchange the compute_metrics with another function I saw, as_dataframe, which doesnt compute a summary, but lists everything individually. If you would like to predict a property like the band gap, which isn't natively supported by MatterSim, I can't really give good advice on that :/ you probably have to train some part of it on your own — Reply to this email directly, view it on GitHub<#160 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/BJK65LCK4JKBA5ILNXNFU233AH37NAVCNFSM6AAAAAB5UJAF4GVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTGMRWGQ4TSNI>. You are receiving this because you authored the thread.Message ID: ***@***.***>