Can universal_Hamiltonian.ckpt be used to predict perturbed structures for subsequent EPC calculations

[3532564593]

Dear Professor Zhong,
Recently, I need to perform EPC (electron-phonon coupling) calculations in my research, which involves the construction and prediction of perturbed structures. Noticing that your team has provided the pre-trained universal_Hamiltonian.ckpt model in the open-source project, I would like to consult you on a technical question:
Can the universal_Hamiltonian.ckpt model be directly used to predict the perturbed structures required for subsequent EPC calculations?
Best regards,
Jiaxin Han

[QuantumLab-ZY]

Dear Jiaxin,

Sure. It is recommended to use the universal model trained based on the 2.0 version of HamGNN.

Best,
Yang

[3532564593]

Thanks for your response. Where is the universal model file for HamGNN2.0?

[QuantumLab-ZY]

You can download the pre-trained model files from the following Zenodo link: https://zenodo.org/records/15568557. Once downloaded, you can load them using Uni-HamGNN. For detailed instructions on how to do this, please refer to the Uni-HamGNN README.

Note that Uni-HamGNN supports predicting both non-SOC (without spin–orbit coupling) and SOC (with spin–orbit coupling) Hamiltonian matrices.

[3532564593]

I have used the universal model (version 2.0) to predict the Hamiltonian, and the Mean Absolute Error (MAE) of the predicted Hamiltonian is very low, which suggests a good overall agreement. However, when I use this Hamiltonian to compute the EPC matrix, the final results differ significantly from those reported in your work.

I understand that EPC calculations can be quite sensitive to subtle details, and I suspect there might be a crucial step or parameter setting in my workflow that I am missing or have misunderstood.

Would you be kind enough to offer some advice or point me towards potential solutions to help resolve this discrepancy? Any insights you could provide would be greatly appreciated.

Thank you very much for your time and consideration.

demo.zip

[QuantumLab-ZY]

The electron-phonon coupling (EPC) value of the longitudinal acoustic branch (b=3) at the Gamma point is divergent. Therefore, the value at this point can be removed.

[3532564593]

I'm sorry, but I didn't quite understand what you meant. My question is that the demo.zip mentioned above is my input file. The first figure shows the EPC matrix for SiC calculated in your paper, while the second figure presents all 6 phonon branches from my calculation—and I found that they do not match at all. Could you please help analyze this issue if you have time, or provide an input file that can reproduce this figure?

[QuantumLab-ZY]

After reviewing your plot, I've noticed that you're currently displaying EPC values for all phonon branches together in a single figure. To properly analyze the data, you should isolate and examine the phonon branch where b = 3 (the longitudinal acoustic branch) specifically. The electron-phonon coupling value for the longitudinal acoustic branch (b=3) at the Gamma point is known to be divergent. This is an expected behavior in the theoretical calculation; therefore, you should exclude the Gamma point data when ploting its figure.