LAMMPS interface for CHGNet

[advancesoftcorp]

Dear developers of CHGNet,

We have developed the interface of LAMMPS for CHGNet: https://github.com/advancesoftcorp/lammps/tree/based-on-lammps_2Jun2022/src/ML-CHGNET
It would be great that it will interest you.

Best Regards,
Satomichi Nishihara
AdvanceSoft Corp.

[bowen-bd]

Thanks a lot!
We appreciate your efforts.
This will be very valuable to users familiar with lammps.

[janosh]

@advancesoftcorp Thank you very much! I assume you've tested the CHGNet interface on an example system? Are you able the share the code for that? We'd be happy to add that to our example notebooks as a starting guide for LAMMPS users.

[advancesoftcorp]

We have tested the CHGNet interface for ZrO2, which input file is here: https://github.com/advancesoftcorp/lammps/tree/based-on-lammps_2Jun2022/examples/CHGNET

[bowen-bd]

@advancesoftcorp
Thanks again!
I see the CHGNet driver file has a input argument for loading different CHGNet weights.
def chgnet_initialize(model_name = None, dftd3 = False, gpu = True):

Say, now we have another CHGNet that can be loaded like myCHGNet = CHGNet.load('new_model')
Could you please share how is it possible to use this model_name key to load different CHGNet in the inp.lammps input file?

[advancesoftcorp]

You can set the model_name at the line of pair_coeff as https://github.com/advancesoftcorp/lammps/blob/22fc3bec34a10c095222b281714af27bedd1677d/examples/CHGNET/inp.lammps#L24C1-L24C36

[advancesoftcorp]

Now, we have already developed the GUI of LAMMPS/CHGNet as the following figure.
Please let us know the name of "new_model", if you have already defined it.
The GUI is our product of Advance/NanoLabo: https://www.nanolabo.advancesoft.jp/en/

[bowen-bd]

@advancesoftcorp
Currently CHGNet.load() method only supports loading the MPtrj pretrained CHGNet.
When users load their own fine-tuned model, it's loaded through CHGNet.from_file('weight_path')
If there can be an interface to load CHGNet from a local path with CHGNet.from_file in the inp.lammps that will be helpful.

[advancesoftcorp]

@BowenD-UCB
I did it.

chgnet_driver.py -> https://github.com/advancesoftcorp/lammps/blob/1ab512916a99e38b46d1d8a6660b0212da95592f/src/ML-CHGNET/chgnet_driver.py#L41C6-L41C6

inp.lammps -> https://github.com/advancesoftcorp/lammps/blob/1ab512916a99e38b46d1d8a6660b0212da95592f/examples/CHGNET/inp.lammps#L25

[bowen-bd]

This looks great, thanks!

[FuXiao-153]

@advancesoftcorp Thank you very much for developing this interface! When I installed the 2JUN2022 version of LAMMPS, some problems occurred. I can run the program using other pair style, but CHGNet fails. I was wondering if the installation of this version is different from the installation of other versions of LAMMPS?

[advancesoftcorp]

Did you make yes-ML-CHGNET, before compile LAMMPS ?

[FuXiao-153]

@advancesoftcorp
yes，before compiling LAMMPS，we completed making yes-[KSPACE/ MANYBODY/ ML-CHGNET/ MOLECULE/ PYTHON/ RIGID].

[advancesoftcorp]

It seems you did not set LD_LIBRARY_PATH for python or compilation was not correct.

[FuXiao-153]

@advancesoftcorp
Sorry to bother you again, we set LD_LIBRARY_PATH for python in .bashrc and have tried different ways of compilation, the program still reported the same error：

LAMMPS (2 Jun 2022)
  using 1 OpenMP thread(s) per MPI task
Reading data file ...
  triclinic box = (0 0 0) to (25.256318 25.844557 25.363354) with tilt (0.997650                                                       81 -1.0715861 1.0101746)
  1 by 1 by 1 MPI processor grid
  reading atoms ...
  1280 atoms
  read_data CPU = 0.036 seconds
Segmentation fault

Could you please help us check the cause of the error? Thank you very much

[advancesoftcorp]

You have to resolve the problem by yourself, because I cannot see what you see.

[mhsiron]

Hello @advancesoftcorp I am using CHGnet plug-in that you have built in LAMMPS mostly successfully. However when the amount of atoms changes such as with a fix dep command (like so):
fix dep0 adatom deposit 100 0 20 7087 region deposregion near 3 mol molec vz -410.8341 -502.1306 vx -0.0 -0.0 vy -0.0 -0.0

I run into the following error due to the number of atoms changing:

ValueError: Array "numbers" has wrong shape (772,) != (768,).
ERROR: Cannot calculate energy, forces and stress by python of CHGNet. (src/ML-CHGNET/pair_chgnet.cpp:672)

I was able to get the code to still work by changing line 92 in chgnet_driver.py from:
global myAtoms
to
myAtoms = None

I'm wondering if there is any downstream effect you suspect from this and why myAtoms was made a global variable?

[advancesoftcorp]

@mhsiron
Thank you for pointing out the bug. As you said, chgnet_driver.py did not support atomic depositions or grand canonical ensembles. But I fix it now:
https://github.com/advancesoftcorp/lammps/blob/d3dc7730a0f2640eb725015e9edaa0b74b222d81/src/ML-CHGNET/chgnet_driver.py#L96

myAtoms is still a global variable, to avoid multiply creating instances of ase.Atoms for speed-up.

[heroiciota]

@advancesoftcorp Hi, When I use your software named Nanalabo pro, which I got licence from your guys days before, to run my own cif structrue, it reminds me
"Traceback (most recent call last):File "C:\Program Files\AdvanceSoft\NanoLabo\chgnet\chgnet_driver.py", line 11, in from chgnet.model import CHGNet, CHGNetCalculator
ModuleNotFoundError: No module named 'chgnet.model'，And I went to log file , there is an ERROR: 'Cannot initialize python for pair_coeff of CHGNet. (../pair_chgnet.cpp:503)' hope to get you guys reply sooner, thx you anyway.

[advancesoftcorp]

@heroiciota
It seems that you did not install chgnet into your machine.
Do "pip install chgnet".

[heroiciota]

@advancesoftcorp Thx！Actually，I have had chgnet installed. But somehow my system can't point to the right environment. I am struggling with it, hopefully I would fix it today.

[heroiciota]

@FuXiao-153 @advancesoftcorp I met the same problem :unrecognized pair style 'chgnet/d3' is part of the ML-CHGNET package which is not enabled in this LAMMPS binary. (../force.cpp:271)Last command: pair_style chgnet/d3 ../../potentials/CHGNET
I wonder did any one succeed in running the example? Can anyone share the experience of it . Also , I have my ML-CHGNET package installed too. Appreciate.

[Mofahdi]

I have an error here:
it seems that
TypeError: chgnet_initialize() takes from 0 to 3 positional arguments but 4 were given

This is what I wrote in my input file which is the same as the tutorial
ERROR: Cannot initialize python for pair_coeff of CHGNet. (src/lammps_releases/advancesoftcorp/lammps/src/ML-CHGNET/pair_chgnet.cpp:519)
Last command: pair_coeff * * 0.3.0 Zr O
How can I solve this issue?

[advancesoftcorp]

@heroiciota
it seems you did NOT make yes-ML-CHGNET.

[advancesoftcorp]

@Mofahdi
the chgnet_driver.py that you had used may be broken.

[naji-usr]

@advancesoftcorp Thank you！If installed in HPC, how can I run parallel calculations (at least is there any way to speed it up!) .

[advancesoftcorp]

@naji-usr you cannot use MPI parallel. but OpenMP is available.

[naji-usr]

Thank you @advancesoftcorp it works.

[ankit213910]

@bowen-bd @advancesoftcorp Hello, I ran CHGNet on LiAlO2. It ran successfully. My goal is to generate a LAMMPS usable potential file. Are you able to help me with that? I see that this link has code that uses a file
pair_coeff * * MPtrj-efsm Zr O
https://github.com/advancesoftcorp/lammps/blob/1ab512916a99e38b46d1d8a6660b0212da95592f/examples/CHGNET/inp.lammps
How to generate the MPtrj-efsm file from the CHGNet code on colab? Also I have lammps Aug 2023 but I dont see the CHGnet module in it.

[advancesoftcorp]

The file of the pretrained model is here:
https://github.com/CederGroupHub/chgnet/tree/main/chgnet/pretrained