IR calculation with Machine Learning Potential

[mind-scraper]

Dear Developers,

Thank you for providing the sscha packages.
I want to calculate the IR spectra of a molecule adsorbed on a metal surface. I use NequIP+LES package (https://github.com/ChengUCB/NequIP-LES/tree/main) that can predict the Born effective charge but not the dielectric tensor. I tried similar steps as in IR Example (https://sscha.eu/Tutorials/tutorial_05_ramanir/), but the code complained that the dielectric tensor is not available.

lanczos = DL.Lanczos(ensemble)
AssertionError: Error, effective charges provided, but not the dielectric tensor.

Disabling the dielectric tensor (lo_to_split=None) gives another error

lanczos = DL.Lanczos(ensemble, lo_to_split=None)
lanczos.ignore_v3 = True
lanczos.ignore_v4 = True

lanczos.init()

polarization = np.array([1,0,0]) # Polarization of light
lanczos.prepare_ir(pol_vec = polarization)

n_iterations = 1000
lanczos.run_FT(n_iterations)
Error, you must initialize a perturbation to start the Lanczos. Use prepare_raman/ir or prepare_perturbation before calling the run method.

Are there any workarounds to fix this?

With best regards,
Samuel

[mesonepigreco]

Yes, you could use an identity dielectric tensor:

import numpy as np

# dyn is the dynamical matrix, i.e., the Phonon() object that also contains the effective charges
dyn.dielectric_tensor = np.eye(3) 

the np.eye(3) builds a 3x3 identity matrix.

In this way, the code goes on without complaining, assuming it is in air. If you have a naive estimation of the material dielectric constant, you can multiply it to the np.eye(3) to get a more realistic IR absorbance.

Another suggestion: you may want to follow the new tutorials for the 2026 SSCHA school, as they are more up-to-date:
school website

There you can find both the PDF with the updated tutorials and the GitHub repository with the code and the source material to run them, if needed. Over the next months, we will replace the main tutorials on the website with these updated ones.

[mind-scraper]

Dear Prof. Monacelli,

I did as you suggested, and the problem is fixed now.
Thank you.

With best regards,
Samuel