The accuracy of MB-pol in high-pressure phases of ice

[oe R]

Dear MBX developers,

Hope this message finds you well!

I've been reading a lot of papers on MB-pol recently, and it does accurately describe the gas–liquid phase of water, as well as behavior in the low to moderate pressure regimes. However, I haven't come across much literature on its performance under high-pressure conditions (10 GPa - 100 GPa). I'm curious how well it holds up in that regime.  

Best regards,

Wenhao

[Xuanyu Zhu]

Dear Wenhao,

As long as water remains molecular, i.e., no proton dissociation occurs, MB-pol *in principle* remain applicable. In practice, however, several factors affect its accuracy, and validation tests are needed:

1. Numerical fitting errors. These could depends on training sets as well as functional forms. An extreme example of numerical error is the "potential energy holes" which can cause simulation instabilities.

Our recent paper on MB-pol(PIFS) explored on this subject. (Just accepted in JCTC, we will update links when it is published. Chemrxiv: https://doi.org/10.26434/chemrxiv-2025-76p6j). Specifically, MB-pol might be numerically unstable above 3GPa (no physical dissociation is expected at this range), and MB-pol(PIFS) are stable. Above ~14GPa, however, ice structures are not fully known experimentally (specifically we care about if dissociation occurs), so it's not necessary that MB-pol(PIFS) should work or not. At higher pressures, ice X are not molecular so the model is not applicable. For liquid, similar idea holds: if water should remain molecular, then MB-pol(PIFS) could work (validations are needed); if it is ionic fluid, then the model is not applicable.

There is this paper that reports some MB-pol lattice energies of high-pressure ice phases (e.g. ice VIII): https://doi.org/10.1021/acs.jctc.6b01248. MB-pol(PIFS) improves a lot on the lattice energy of ice VIII (see Supporting Information of the MB-pol(PIFS) paper after link is updated).

Here's a temporary github to MB-pol(PIFS). (Note: under "PIFS" branch) https://github.com/PhilipZhu/MBX/tree/pifs.

2. MB-pol uses machine-learned potential upto 3-body energies, and uses classical polarizable model for 4+ bodies; The error coming from 4+ body energies in lower pressures (<1GPa) seems small, but it remains to be tested for higher pressures.

We are currently investigating the behavior of ice VII.

XZ

[oe R]

Hi Xuanyu,

Thank you for your prompt response! 

Is the Python interface to MB-pol (PIFS) still applicable? 

Best,

Wenhao

[Xuanyu Zhu]

Yes, it should be. Feel free to let us know if you encounter any problems.

Here's the updated Chemrxiv link to the latest accepted version PIFS paper that we mentioned in the last message: https://doi.org/10.26434/chemrxiv-2025-76p6j-v2