Add MLIP Tutorial

[PythonFZ]

Add a Tutorial on Machine-Learning Interatomic Potentials

This tutorial requires some additional tools:

• https://github.com/ACEsuit/mace
• https://github.com/zincware/rdkit2ase
• https://github.com/m3g/packmol (^20.15.0)
• https://github.com/zincware/ZnDraw/releases/tag/v0.4.7 or newer

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[RudolfWeeber]

Thank you, Fabian.
@SamTov agreed to do a detailed review next week.

Whlie having a quick look, I noticed the following:

• there are some terms and abbreviaitons which people new to the field might not know. A bit more contet and explanaiton in the intro would be useful
• Maybe the role of the different external parckges could be stated mor explicilty. In particular, how Espresso, ASe and Mace interact
• Units: It wans't completely clear to me at first glance what units are used in the simulatoion. Is ESPResSo run in the usual reduced units?
• As long as there are only external forces, the Verlet-Lists and skin can//should eb switche off
• It would be important to get the external depencies stable, as we will have to include them in the CI

[PythonFZ]

Hey Rudolf,

Thanks for the feedback. I have added some more clarification w.r.t. to the code that is used and added some more sources. I'd appreciate if you could help me improve which parts still need more clarification.

• Units: It wans't completely clear to me at first glance what units are used in the simulatoion. Is ESPResSo run in the usual reduced units?

I have talked to @reinaual about the units and if I understood it correctly, I am defining the unit system in ESPResSo such that it resembles the one used within ASE.

• As long as there are only external forces, the Verlet-Lists and skin can//should eb switche off

As I am not an ESPResSo user, would it be possible to give me an example or edit the file direclty.

• It would be important to get the external depencies stable, as we will have to include them in the CI

I have removed the ipsuite dependency because all the features required are within the rdkit2ase package.
The new dependencies would be:

• pip install mace-torch==0.3.6
• pip install rdkit2ase==0.1.4
• https://github.com/m3g/packmol/releases/tag/v20.15.1

[SamTov]

• I think the title should either be Machine-Learned Interatomic Potentials or Machine Learning for Interatomic Potentials.
• Maybe mention what kind model is? Something like, Equivariant graph model would be sufficient along with a link to the model paper. This is just the foundation model paper but a direct reference to the architecture would be nice too.
• Do you have an overview of the ASE features somewhere?

Reply via ReviewNB

[SamTov]

Why not explain each import to highlight what it is and why you need it? You can start with a code block and then write out a list of why you need each one. For example, pandas is not a clear thing to import in a simulation study unless you need it somewhere. I also usually split my imports by category in tutorials, something like:

# Simulation imports
import espressomd
from espressomd.plugins.ase import ASEInterface
from mace.calculators import mace_mp
import pint

# Linalg
import numpy as np
import pandas as pd

# Plotting and Visualisation
import plotly.express as px
from espressomd.zn import Visualizer

# Utils
import tqdm

# Analysis
from rdkit2ase import smiles2atoms, pack

Reply via ReviewNB

[SamTov]

You can use full variable names to alleviate any uncertainty. Things aren't clear to everyone. For me, kb would be clearer than boltmk but as you never introduce the fact that you will talk about Boltzmann, I would recommend just writing out boltzmann_constant or boltzmann_k. Maybe also add comments after each line highlight what unit conversions you are doing.

Reply via ReviewNB

[SamTov]

This is pretty random after you talk about SMILES strings. Why not have a section about what this initialisation does and then show it. After, introduce SMILES and do the system construction.

Reply via ReviewNB

[SamTov]

The peopl at the summer school may not know what SMILES is. For the sake of understanding what "CCO" means, it is probably worth just putting a picture of ethanol in and showing how the SMILES string is built out of it.

Reply via ReviewNB

[SamTov]

Line #1.    for atom in etoh:

It isn't clear why the atoms should be in the model object. Have you introduced the type mapping? What it is, why you need it?

Reply via ReviewNB

[SamTov]

Should you tell people to open it in a different window as they will be running the descent after the opening? Maybe go to split screen or something?

Reply via ReviewNB

[SamTov]

• We haven't seen the first parts yet. Up to now, I don't know if it has accurate forces or can run MD. There is no comparison. You can do a classical force-field comparison, for example, or show DFT data. I actually have DFT ethanol trajectories you can use. Or, run a short DFT ethanol MD with PySCF, it is pretty fast.
• Maybe list some chemical reaction methods.

Reply via ReviewNB

[SamTov]

Line #3.    system.thermostat.turn_off()

Why? People will ask why you are doing things like this. If you had to ask someone for help in making the script, the people in the tutorial will be at the same level but also without the ML knowledge. Explain everything.

Reply via ReviewNB

[SamTov]

Write up a conlcusion, what have we learned, what can we do next, why is this cool.

Reply via ReviewNB

[SamTov]

Let us consider what we want from the tutorial. Is a reaction the only analysis we want to do? Don't most people want to run MD before doing a reaction study? Both are nice, but we could do RDF calculations or even Rg computations on something like ethanol. Or go for a bigger molecular and do a quick polymer study.

[PythonFZ]

Is there any reason you removed the interactive plotting part for the visualizer vis.zndraw.figures = {"distance": fig.to_json()}?

[SamTov]

Is there any reason you removed the interactive plotting part for the visualizer vis.zndraw.figures = {"distance": fig.to_json()}?

It didn't work when we ran it on the CIP pool computers.

[PythonFZ]

Is there any reason you removed the interactive plotting part for the visualizer vis.zndraw.figures = {"distance": fig.to_json()}?

It didn't work when we ran it on the CIP pool computers.

Is there any reason you removed the interactive plotting part for the visualizer vis.zndraw.figures = {"distance": fig.to_json()}?

It didn't work when we ran it on the CIP pool computers.

I see - I opened #4993 which would resolve this and make it work.

[SamTov]

Is there any reason you removed the interactive plotting part for the visualizer vis.zndraw.figures = {"distance": fig.to_json()}?

It didn't work when we ran it on the CIP pool computers.

Is there any reason you removed the interactive plotting part for the visualizer vis.zndraw.figures = {"distance": fig.to_json()}?

It didn't work when we ran it on the CIP pool computers.

I see - I opened #4993 which would resolve this and make it work.

Is it necessary? They can plot the plotly figure in the notebook and the syntax to have it opened in ZnDraw is quite odd, particularly for those who aren't familiar with the program.

[PythonFZ]

Is there any reason you removed the interactive plotting part for the visualizer vis.zndraw.figures = {"distance": fig.to_json()}?

It didn't work when we ran it on the CIP pool computers.

Is there any reason you removed the interactive plotting part for the visualizer vis.zndraw.figures = {"distance": fig.to_json()}?

It didn't work when we ran it on the CIP pool computers.

I see - I opened #4993 which would resolve this and make it work.

Is it necessary? They can plot the plotly figure in the notebook and the syntax to have it opened in ZnDraw is quite odd, particularly for those who aren't familiar with the program.

• opened in ZnDraw is quite odd could you elaborate what is odd about it and ideally give a suggestion on how to improve it?
• I would like to showcase the interactive part of having plots alongside the data in the visualizer. It is not required but I thought it is a nice touch. From the MLIP standpoint it is not required.

[SamTov]

I like it.

[PythonFZ]

Thanks for all the work you put into this 👍

The

#TODO: figure out if this works

as well as the vis.zndraw.figures = {"distance": fig.to_json()} might both depend on #4993

Unfortunately, I don't know what issues came up with this version and they did not mention or make any issues on the ZnDraw repo. If the issue will not be resolved before the tutorial it might be easiest to remove this part.

[jngrad]

• this energy value needs a citation
• this energy value is not reproduced by the MACE-MP-0 model in the current form of the tutorial

[SamTov]

We will remove this as we have no good explanation for why it does not come out right. It could be due to the temperature or a spurious comparison value.

[SamTov]

Energy is now resolved and correct.

[jngrad]

It's unclear to me, what is the answer to this question. The dip in energy could be due to the exothermic reaction $\mathrm{H_2SO_4 + H_2O \to HSO_4^{-} + H_3O^{+}}$, but the energy difference doesn't reproduce experimentally determined Gibbs free energy values. I'm also unsure if the potential energy can be directly equated to a Gibbs free energy, a Helmholtz free energy, an enthalpy or an internal energy.

[SamTov]

I would actually like to discuss this point during the tutorial. The question of whether this is a physically realistic process is fundamental to the explainability of ML potentials. During the discussion, I would also ask how to extend the study to determine the answer. Namely, running it longer and generating some statistics or pH measurements that can be compared to experimental values.

[RudolfWeeber]

Is something till missing or can ew go ahed?

[SamTov]

The ground state energies for the water study are resolved and correct.