20 Jun 2024 |
Adam Jackson | In general the difference between a "phonon calculation" and "vibrations calculation" is that with "phonons" one deals with reciprocal-space matters (either by DFPT sampling of q-point meshes or by supercell methods). If we don't care about phonon wavevectors and just want information about the simulation region, the implementation can be a lot simpler. | 10:49:07 |
Adam Jackson | So in ASE "Vibrations" can actually be used with periodic boundary conditions if you like, and will just give information about Gamma-point. (This is quite useful for catalysis etc.) But it never makes sense to use "Phonons" with open boundary conditions | 10:50:26 |
Adam Jackson | (Admittedly this overlooks 2D or 1D periodic conditions... but so do most implementations) | 10:51:05 |
Ask Hjorth Larsen | Bar is open over in #ase-bar:matrix.org | 18:59:19 |
Hasith Vattikuti | Thanks for the advice, the changes actually seemed to give me the 2F in the right spots, the only difference is that F now has significant non-zero elements along the diagonal which means that E_yy and E_zz are somehow nonzero. Would anyone know what caused this even though all the atoms all lie along the x axis? | 20:16:46 |
Hasith Vattikuti | The Hessian in question
[[-122.9 0. 0. 61.6 0. 0. 61.6 0. 0. ]
[ 0. -146.6 0. 0. 73.4 0. 0. 73.4 0. ]
[ 0. 0. -146.6 0. 0. 73.4 0. 0. 73.4]
[ 61.6 0. 0. -122.9 0. 0. 61.6 0. 0. ]
[ 0. 73.4 0. 0. -146.6 0. 0. 73.4 0. ]
[ 0. 0. 73.4 0. 0. -146.6 0. 0. 73.4]
[ 61.6 0. 0. 61.6 0. 0. -122.9 0. 0. ]
[ 0. 73.4 0. 0. 73.4 0. 0. -146.6 0. ]
[ 0. 0. 73.4 0. 0. 73.4 0. 0. -146.6]]
| 20:28:07 |
Hasith Vattikuti | Oh it seems like Vibrations will respect fixed constraints, maybe I should try fixing the atoms to only move along the x axis. But it is still curious as to why E_yy and E_zz are non-zer | 20:36:44 |
Hasith Vattikuti | * Oh it seems like Vibrations will respect fixed constraints, maybe I should try fixing the atoms to only move along the x axis. But it is still curious as to why E_yy and E_zz are non-zero | 20:36:59 |
Adam Jackson | Do you not expect to get transverse vibrations? | 20:37:18 |
Adam Jackson | I'd expect a restoring force when moving atoms out-of-line from the chain | 20:42:09 |
Hasith Vattikuti | Oh true, I guess it slipped my mind that it would be almost the same case in the y and z dimensions | 21:17:52 |
25 Jun 2024 |
Robert Warmbier | Does anyone here have experience generating MD potentials from DFT data? I'm trying to get into it. Lots of people produce FFs (like the reaxFF ones), but rarely any software or toolkit is mentioned. Is there anything user-friendly out there? | 14:17:58 |
Anubhab Haldar | Classical potential or MLIP? | 14:18:47 |
Robert Warmbier | Hm, does it make a difference? | 14:19:14 |
Robert Warmbier | In the sense of final result. | 14:19:38 |
Robert Warmbier | Something like reaxFF is classical per se, but still people use machine learning to get it. | 14:20:43 |
Robert Warmbier | ah, MLIP means no functional form, not (just) the way it's optimised... | 14:21:51 |
Robert Warmbier | my bad | 14:21:56 |
Robert Warmbier | Anything which can produce a FF for a disordered material with different bond orders does the job. | 14:23:15 |
TomD | There is a few papers available, MC followed by active learning seem to be popular | 14:31:31 |
mjwaters | In reply to @rbw:matrix.org Does anyone here have experience generating MD potentials from DFT data? I'm trying to get into it. Lots of people produce FFs (like the reaxFF ones), but rarely any software or toolkit is mentioned. Is there anything user-friendly out there? I think pacemaker is very friendly. | 15:39:04 |
mjwaters | but the secret sauce is always what structures do you train on? | 15:39:50 |
mjwaters | In reply to @rbw:matrix.org Does anyone here have experience generating MD potentials from DFT data? I'm trying to get into it. Lots of people produce FFs (like the reaxFF ones), but rarely any software or toolkit is mentioned. Is there anything user-friendly out there? there are a few of us that train MLIPs here. | 15:41:07 |
26 Jun 2024 |
Toma Susi | I honestly wouldn't bother training a classical potential for a specific system in 2024 instead of an ML potential, unless you have a very good reason to | 07:25:30 |
Toma Susi | The former is probably harder to do well, and very likely will result in poorer accuracy for your system | 07:26:03 |
Hasith Vattikuti | Is there a general rule of thumb for when one should switch from DFT to MLIPs? For example is 50 carbon atoms too computationally heavy for DFT compared to MLIPs? What about 100? | 15:33:51 |
mjwaters | In reply to @hasithv:matrix.org Is there a general rule of thumb for when one should switch from DFT to MLIPs? For example is 50 carbon atoms too computationally heavy for DFT compared to MLIPs? What about 100? Depends on how many times you will evaluate forces and energies. | 17:44:41 |
Hasith Vattikuti | Lets say I'd evaluate forces and energies both on the order of O(n^2), n=atoms, which sounds reasonable for a good amount of applications I believe | 18:29:55 |
mjwaters | and DFT costs roughly O(n^3) | 18:46:46 |
mjwaters | so n^5 | 18:46:57 |