I’ve been looking recently into the fitting of oxides potential parameters (including a core-shell representation) using phonon frequencies as observables to try to reproduce complex dielectric spectra. With the help of the documentation and several topics on this very forum, I’ve managed to have the fitting process running, using either experimental or DFT data.
Now my question is related to a remark I read in one of your paper, about the derivation of potential parameters of aluminophosphates and berlinite (J. Chem. Soc., Faraday Trans., 1994, 90, 3175-3179), namely: “The phonon frequencies were excluded from the fitting procedure as this requires the correct assignment of all modes to their corresponding experimental frequencies at all stages of the optimisation of the potential parameters”. Upon reading that, and considering I’m quite new to GULP, I’m not sure how to what extent I can trust my procedure. Since some time have passed since this paper and its contemporary version of GULP, what’s your take on that point now?

Dear Antoine,
The issue still remains. While you can fit the frequencies in GULP the problem comes if the modes are in the wrong order - you have no guarantee that the fit will be making the right mode have the right frequency. If you have a simple system where the modes are well separated & a good starting point that already has the modes in the correct order then it might work. What is really needed is to input the eigenvector and frequency so that the code can look for the maximum overlap of the input mode with the computed one. Unfortunately this isn’t implemented because the input would get more complex for the user and it would be essential to make sure the same eigenvector convention is being used (mass vs non-mass weighted etc).
One solution is to fit the curvature of the energy surface by giving a series of energies vs geometries as this is related to the phonons, but is unambiguous.
Hope that helps.
Best regards,
Julian

Hi Antoine,
I was actually having a bit of senior moment yesterday. I thought I might have added the fitting to eigenvector / frequency combinations, but initially couldn’t find it. Now I’ve discovered that it is implemented after all. If you look at the “mode” option in the help text this will hopefully explain.
Sorry for the incorrect initial information, but at least it’s a good news outcome.
Best regards,
Julian

haha many thanks for this clarification. I did see the example with the water molecule using this observable, but I wasn’t sure about its relevance. Now it’s clearer and I’ll use the eigenvector/frequency sets from DFTPT calculations then.
Good news indeed, thanks again!!

I have a follow-up question regarding the fitting of potential parameters using phonon frequencies, especially at the gamma point. Is the non-analytic correction taken into account for the fitting, for example when specifying “gamma_direction_of_approach” or asking for dispersion curves?

Also, another related question, the calculation of the dielectric complex function never uses phonon frequencies calculated with the non-analytic correction, even when GULP is given a “gamma_direction_of_approach”, right?

Dear Antoine,
Normally the non-analytic correction is only included in the frequencies if the k point is part of a dispersion curve, in which case the direction comes from the dispersion direction specified. If you use “gamma_direction_of_approach” then the non-analytic correction should be included even for a single k point and so this would change the frequencies for fitting at gamma.
The frequency-dependent properties are computed using the gamma point frequencies. This means can probably force the use of a non-analytic correction at gamma using the “gamma_direction” option and this will get passed through.
Best regards
Julian

thanks again for your enlightening answer. Regarding the frequency-dependent properties, invoking the “gamma_direction” option, or the “dispersion” one, doesn’t influence them, while still showing the TO/LO splitting. The same dielectric spectrum is produced whether or not the non-analytical correction is used and I’m a bit confused by that since some gamma-point frequencies (especially a few involved in dielectric peaks) are different. I tried that on several systems, including the MgO used in example 5 of GULP. I’m still a bit new to phonon-related studies so is this normal and expected?

Dear Antoine,
I’ve just checked the code and it seems that the frequency-dependent properties currently only use the frequencies without LO-TO splitting, so apologies for giving you false hope. I can look at changing this for the next release. If you feeling like having some fun I think you can probably change this in the code. If you go to phonon.F90 (assuming you are doing something standard on a single processor) and look for the call to omegaproperty. If you change the argument “savefreq” to “freq” then you may get what you are looking for.
Best regards,
Julian

Thanks again! Indeed, this modification works and I get the modified frequency-dependent properties! Now, upon reading about LO-TO splitting, I’m not sure if that entirely makes sense to take it into account when considering only the complex dielectric function at the gamma point precisely. I understand this splitting occurring near k=0, but not really exactly at k=0, since there’s no direction of propagation per say in itself (or it doesn’t make sense to just think about the gamma point just by itself). But this is not related to GULP haha.
Thanks again for answering all my GULP-related questions!

Dear Antoine,
The debate over whether a directional contribution should be included due to LO/TO is probably why this wasn’t included in the term is probably why it was excluded originally. Of course the measurement of frequency-dependent properties as an in and out bound direction for the wave (given by odirection) and so arguably any splitting may reflect this, but I’ve yet to find a definitive reference that addresses this.
Best regards,
Julian

I had a follow-up question related to this discussion. If I want to add phonon frequencies from IR measurements as an observable, would I be correct in thinking ‘gamma_angular_steps’ should be set to some nonzero value, to account for the LO-TO splitting?

Hi Connor
If you really want to handle the LO-TO splitting in detail then in principle you should average over the directions of approach. However, fitting these modes at that level of detail would be tricky since they are delocalised, unlike the fingerprint high frequency modes that people usually get from an IR spectrum. If your objective is really to nail LO-TO splitting then you’ll probably need to fit phonon dispersion curves rather than just look at the gamma point.
Best regards,
Julian