Several crystal phases from the same trained potential (polymorphism)

[mfhm]

Hi,

You have done a great job with deePMD code! Is it possible to train a potential, from which one could get different crystal phases depending on temperature and/or pressure? Are you aware of any publications, which have successfully demonstrated this?

Thanks in advance,
Max

[luigibonati]

Dear Max, Yes, it is possible. You should take care of including not only configurations from the bulk structures but also interfaces between them as well as defects. There are many approaches for this task. You can collect the configurations just from plain MD, you can use active learning to select new configurations based on ensemble uncertainty (see DeepMD-GEN) or you can use enhanced sampling techniques to efficiently collect all the relevant structures. An example of the latter is the following: https://journals.aps.org/prl/abstract/10.1103/PhysRevLett.121.265701 . In this case we have done it for the liquid and the cubic diamond structure of silicon but this can be extended using the same procedure to larger portion of phase diagrams. We have an another example which it's out on review, I hope it will be out soon! Best, Luigi

…

On Tue, Sep 24, 2019 at 8:13 AM mfhm ***@***.***> wrote: Hi, You have done a great job with deePMD code! Is it possible to train a potential, from which one could get different crystal phases depending on temperature and/or pressure? Are you aware of any publications, which have successfully demonstrated this? Thanks in advance, Max — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <https://github.com/deepmodeling/deepmd-kit/issues/74?email_source=notifications&email_token=ACRCUS4BIE5CJ2YXZZL3H53QLIVAHA5CNFSM4I2BBXX2YY3PNVWWK3TUL52HS4DFUVEXG43VMWVGG33NNVSW45C7NFSM4HNLD27Q>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ACRCUS74YYEPLO56HWXCFDDQLIVAHANCNFSM4I2BBXXQ> .

[jameswind]

Thank you so much, Luigi, for your reply! Hi Max, I confirm that the answer is yes and we have seen many successful examples. Many of them are still under extensive tests, so my suggestion at this moment is to take a look at Luigi's work on silicon and the active learning approach ( https://journals.aps.org/prmaterials/abstract/10.1103/PhysRevMaterials.3.023804). The latter realizes a PES model that is reliable for a large configuration space, including several crystalline structures and liquids, for the Al-Mg system. The approach is not limited to metallic systems. Please see the Github repo dpgen (https://github.com/deepmodeling/dpgen) for more information. dpgen is still under active development and any feedback or contribution of yours would be greatly appreciated. Best, Linfeng On Tue, Sep 24, 2019 at 12:33 PM Luigi Bonati <[email protected]> wrote:

…

Dear Max, Yes, it is possible. You should take care of including not only configurations from the bulk structures but also interfaces between them as well as defects. There are many approaches for this task. You can collect the configurations just from plain MD, you can use active learning to select new configurations based on ensemble uncertainty (see DeepMD-GEN) or you can use enhanced sampling techniques to efficiently collect all the relevant structures. An example of the latter is the following: https://journals.aps.org/prl/abstract/10.1103/PhysRevLett.121.265701 . In this case we have done it for the liquid and the cubic diamond structure of silicon but this can be extended using the same procedure to larger portion of phase diagrams. We have an another example which it's out on review, I hope it will be out soon! Best, Luigi On Tue, Sep 24, 2019 at 8:13 AM mfhm ***@***.***> wrote: > Hi, > > You have done a great job with deePMD code! Is it possible to train a > potential, from which one could get different crystal phases depending on > temperature and/or pressure? Are you aware of any publications, which have > successfully demonstrated this? > > Thanks in advance, > Max > > — > You are receiving this because you are subscribed to this thread. > Reply to this email directly, view it on GitHub > < https://github.com/deepmodeling/deepmd-kit/issues/74?email_source=notifications&email_token=ACRCUS4BIE5CJ2YXZZL3H53QLIVAHA5CNFSM4I2BBXX2YY3PNVWWK3TUL52HS4DFUVEXG43VMWVGG33NNVSW45C7NFSM4HNLD27Q >, > or mute the thread > < https://github.com/notifications/unsubscribe-auth/ACRCUS74YYEPLO56HWXCFDDQLIVAHANCNFSM4I2BBXXQ > > . > — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <https://github.com/deepmodeling/deepmd-kit/issues/74?email_source=notifications&email_token=AEJ6DC3FAS4KS4DDJ6ELQADQLI6L7A5CNFSM4I2BBXX2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD7O7LKI#issuecomment-534640041>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AEJ6DCZN454J7TZUAQ4AHQTQLI6L7ANCNFSM4I2BBXXQ> .