2025-02-13 Force Field Release Meeting notes

Evaluator update

LW

(slide 6)

JW: Can you walk me through the x/y axis?

LW: X axis is the viscosity of predominate component, the y axis is the relaxation time.

JW: So you expect a diagonal line where more viscous take longer to equilibrate

LW: Sort of, but it doesn’t always work that way.

JW: So what does that cluster at the bottom mean?

LW: Those have a higher mole fraction of water

CB – Units of viscosity here are “pascal seconds” - Is there a direct/intuitive interpretation of this? Like “difficulty of stirring”? In particular if there was some way to estimate viscosity from chemical structure this may help evaluator understand how much sampling to do/which settings to use.

MT – Did myh phd on ionic liquids (which are very viscous). It sounds like CB is thinking about viscosity as a structural property, but it’s actually a transport property. My expeirence was that when FFs get things wrong, the viscosity is very badly overestimated since it makes molecles “stick together” too much due to electrostatics or other things. But generally I’d encourage you to think of it as a transport property like diffusivity, not a structural property.

CB – So if we think of this as diffusion instead of viscosity, is that something that could turn into an automated knob/setting/parameter? Like, we tell evaluator what to expect for diffusivity to help it more accurately determine when things are equilibrated? This may be a change we can make to the protocol itself that will give us more confidence in our calculation results.

LW – Good idea, IF we were determined that we wanted to include triethanolamine in the training set. I’d argue that it doesn’t bring enugh unique value to justify its computational cost (especially since it will hold up the entire iteration, even if the other components finish first)

CB – So, we want to think about this is two segments: 1) do we have an issue with viscosity in general and 2) do we want to exclude triethanolamine specifically to avoid a face of the viscosity issue?

LW – The second one is something we can probe experimentally - drop triethanolamine from the training set and see what happens.

JR – So CB, are you thinking of using the diffusivity of a compound as a proxy for its viscosity? … Can anything be gleaned by looking at the interaction energy of the viscous component with itself vs. with water?

MT – I don’t know off the top of my head. I’m kinda in line with LW’s proposal to drop triethanolamine.

(slide 9)

JC – How many mols per box?

LW – 1000

JC – That’s a good size…. thought still wondering if we’re getting weird collective effects like water structure that might be fixed by a larger box.

LW – I’d thought a little about that. I chose 1k for speed, but it’s worth thinking about going larger. Also worth considering increased sampling.

(slide 10)

DM – Would this suggest that this is a significant enough bug that we shouldn’t consider the other results until it’s fixed?

LW – Yes.

JW – So this is just different results from boxes of water, before being mixed with other compoentns? And the enthalpies are consistent?

LW – Yes

MT – Any idea where to start on this?

LW – No special insight, I was going to start from the trajectories and work my way back.

CB – What’s the time autocorrelation of enthalpy for a box of 1k waters?

LW – We see water converge quite quickly. (shows slide 12)

CB – These look like very short term fluctuations (so that’s some evidence against limited box size/finite size effects?)

JC – Are you running in NPT ensemble?

LW – Yes

(See recording ~32 mins for below)

JC – My understanding is that transport properties should be evaluated in NV?T… Wondering if you’re getting finite size effects. Would recommend running in NVT … (JC had a good idea here but may need to fill in the chain of experiments she proposed)

DM – LiveCOMS paper recommends running NVE for transport properties.

JC – Right, that’s because variations in box size … (JC may need to fill this in)

DM – …

LW – Ok, but I’m fairly convinced that there’s a bug in Evaluator that’s before all this.

PB (chat) – Is there a resolution to the packmol issues raised before related to equilibration times? BMorales had presented on one.

LW – I think the splitting of the equilibration and pre-equilibration was aimed at resolving this.

PB – I wonder if this two-stage approach is leading to consistently different/bad starting points.

MT – What if we made boxes bigger and ran stuff for longer?

LW – More sampling did improve things. But I’d want this to be happening in pre-equilibration phase.

CB – Box shape? Increasing surface area to volume ratio may more thoroughly.

JC – Important to have box dims equal in all dimensions to ensure you can average them.

…

MT – Re slide 6 – Do you believe this pymbar numbers for more viscous mixtures? I don’t know much about this, and wouldn’t be surprised if there are assumptions in the equilibration criteria that don’t suit all systems.

LW – I’m not sure that pymbar is great at catching longer timescale movements. So I think this is an underestimate, since I only simulated 12ns. For the fast-equilibrating boxes pymbar gives a reasonable estimate, but for other I wouldn’t trust it.

CB – So are you thinking of removing triethanolamine from training set?

LW – I’ll plan to find some other properties involving trieathanolamine and see if they also have bad estimates due to (what we’re thinking is) the high viscosity.