!JEiuNJLuxedbohAOuH:matrix.org

ASE

185 Members
Atomic Simulation Environment — https://wiki.fysik.dtu.dk/ase/ — Channel for discussions, support, and development of ASE4 Servers

Load older messages


SenderMessageTime
1 Mar 2024
@ajjackson:matrix.orgAdam Jacksonbut hey it ran in 1 second so21:41:37
@ajjackson:matrix.orgAdam JacksonOk I am going to kill this container full of ML dependencies now, this is the last chance to request any re-runs 😆21:43:43
@chronum:matrix.orgAnubhab HaldarContainer purge best purge.21:54:06
@ajjackson:matrix.orgAdam Jackson *
atoms = ase.build.molecule("H2O"
atoms.calc = mace_mp(model="medium", dispersion=False, default_dtype="float64", device="cpu")

opt = LBFGS(atoms)
opt.run(fmax=1e-8)
vib = Vibrations(atoms, delta=0.001, nfree=4)
vib.run()
vib.summary()

gives me

---------------------
  #    meV     cm^-1
---------------------
  0    0.0i      0.2i
  1    0.0       0.0
  2    0.0       0.0
  3    0.0       0.0
  4    0.0       0.0
  5    0.0       0.0
  6  198.5    1601.1
  7  445.4    3592.3
  8  463.3    3736.6
---------------------
Zero-point energy: 0.554 eV
22:06:01
2 Mar 2024
@potatoboiler:matrix.orgpotatoboiler joined the room.11:38:08
@marianomollo:matrix.orgmariano
In reply to @mjwaters:matrix.org
are those imaginary frequencies the translational and rotational modes of the molecule?
yes they are the transformations on the whole molecule
12:31:58
@marianomollo:matrix.orgmariano
In reply to @ajjackson:matrix.org

A lot of things this could be

  • force tolerance of the optimiser needs to be very tight
  • even then the finite difference scheme is unlikely to get the trans/rot modes to exactly zero, usually one aims for "small enough" values which are often negative
  • a "noisy" potential energy surface is more likely to have problems with consistency between a local minimum and finite-difference vibrations. With ML potentials this can be an issue especially if using single-precision GPU evaluation
  • on the other hand, in principle a well-implemented ML potential can be regularised to a nice smooth surface and may behave "better" than the underlying DFT calcs...
i will try tightening the force tolerances more, especially on the small and large models; MACE medium converges for me too
12:33:37
@marianomollo:matrix.orgmariano
In reply to @ajjackson:matrix.org

A lot of things this could be

  • force tolerance of the optimiser needs to be very tight
  • even then the finite difference scheme is unlikely to get the trans/rot modes to exactly zero, usually one aims for "small enough" values which are often negative
  • a "noisy" potential energy surface is more likely to have problems with consistency between a local minimum and finite-difference vibrations. With ML potentials this can be an issue especially if using single-precision GPU evaluation
  • on the other hand, in principle a well-implemented ML potential can be regularised to a nice smooth surface and may behave "better" than the underlying DFT calcs...
i might be interested also in how one can regularise the potential, i have no clue at the moment
12:34:17
@marianomollo:matrix.orgmariano
In reply to @ajjackson:matrix.org
I just tried it and got pretty clean results by using float64, a very tight optimisation and very small displacements
thank you for taking time to reproduce, it really helps!
12:34:57
@ajjackson:matrix.orgAdam JacksonI expect for MACE-MP-0 this has already been done pretty well; the paper should mention it if so. And some of the authors use this room so maybe they can confirm...12:47:04
@ajjackson:matrix.orgAdam Jackson

Question from the MatSci forum that maybe the MD experts can help with: https://matsci.org/t/seeking-help-restarting-molecular-dynamics-simulations-in-ase/53652
In short the question is "how do a restart an MD run that hit walltime?"

They are using Langevin. I think this thermostat has no "memory" of previous steps, so the answer is simply "restart using the velocities from the last step". Is that correct?

14:01:52
@ajjackson:matrix.orgAdam JacksonThe results will not be identical to a longer run with no restart, unless somehow the state of the random-number-generator is also dumped and restarted. But they should be equally valid?14:02:44
@ajjackson:matrix.orgAdam Jackson *

Question from the MatSci forum that maybe the MD experts can help with: https://matsci.org/t/seeking-help-restarting-molecular-dynamics-simulations-in-ase/53652
In short the question is "how do a restart an MD run that hit walltime?"

They are using a Langevin thermostat. I think this one has no "memory" of previous steps, so the answer is simply "restart using the velocities from the last step". Is that correct?

14:03:18
@tomdmre:matrix.orgTomDI would say that if the random number generator is carried over as well with the momenta the restart should lead to similar results?14:20:55
3 Mar 2024
@mjwaters:matrix.orgmjwaterswhich MD method takes an RNG? I thought that was just for initialization of velocities?04:28:17
@chronum:matrix.orgAnubhab HaldarI thought everything N*T in ASE used RNGs?04:45:52
@chronum:matrix.orgAnubhab HaldarSince NVE is deterministic(ish).04:45:59
@mjwaters:matrix.orgmjwatersin N(P,V)T velocity and momentum re-scaling is deterministic right? so I'm not sure where rngs come in other than initialization. 16:02:09
@chronum:matrix.orgAnubhab HaldarDon't thermostats need an RNG? I know there are some deterministic thermostats/dynamics/couplings, but I know Andersen uses an RNG.16:08:05
@chronum:matrix.orgAnubhab HaldarNot as sure about NHC.16:08:12
@mjwaters:matrix.orgmjwaterslooks like we need an MD theorist. 16:09:16
@tomdmre:matrix.orgTomDLangevin will need RNG, normal function I think16:09:20
@tomdmre:matrix.orgTomDLangevin use random collision effects, which viscosity is set using the friction parameter16:11:07
@flokno:matrix.orgFlorian KnoopMost thermostat use random fluctuations, but Nose-Hoover is deterministic: https://en.wikipedia.org/wiki/Nos%C3%A9%E2%80%93Hoover_thermostat17:05:40
@flokno:matrix.orgFlorian Knoop
In reply to @ajjackson:matrix.org

Question from the MatSci forum that maybe the MD experts can help with: https://matsci.org/t/seeking-help-restarting-molecular-dynamics-simulations-in-ase/53652
In short the question is "how do a restart an MD run that hit walltime?"

They are using a Langevin thermostat. I think this one has no "memory" of previous steps, so the answer is simply "restart using the velocities from the last step". Is that correct?

that is correct, yes. as a matte of fact I answered the question with code examples in the ASE issue for presumably the same problem: https://gitlab.com/ase/ase/-/issues/1439#note_1780548998
17:07:45
@flokno:matrix.orgFlorian Knoop
In reply to @ajjackson:matrix.org

Question from the MatSci forum that maybe the MD experts can help with: https://matsci.org/t/seeking-help-restarting-molecular-dynamics-simulations-in-ase/53652
In short the question is "how do a restart an MD run that hit walltime?"

They are using a Langevin thermostat. I think this one has no "memory" of previous steps, so the answer is simply "restart using the velocities from the last step". Is that correct?

* that is correct, yes. as a matter of fact I answered the question with code examples in the ASE issue for presumably the same problem: https://gitlab.com/ase/ase/-/issues/1439#note_1780548998
17:07:54
@flokno:matrix.orgFlorian Knoopthat is, the technical aspect of it17:14:12
4 Mar 2024
@mjwaters:matrix.orgmjwatersI was making a MEP animation and realized that my own slacking on my MRs is biting me. 04:19:59
@mjwaters:matrix.orgmjwatersI need to finish that plotting variables overhaul ASAP. 04:20:15
@elichad:matrix.org@elichad:matrix.org left the room.16:25:37

There are no newer messages yet.


Back to Room ListRoom Version: 5