Using DeePMD MLIP (model.pth/model.pb) with eOn

[Mem314]

Hi eOn developers,

I’m trying to use a DeePMD machine-learned interatomic potential (DeePMD-kit) with eOn for aKMC/rare-event searches.

Context:

I installed eOn 2.11.0 from conda-forge.

My DeePMD potential is available as either:

TensorFlow frozen model: model.pb, or
PyTorch frozen model: model.pth (DeePMD v3 backend).

I understand eOn can use LAMMPS as a potential interface, but my understanding is that eOn’s LAMMPS interface is typically set up with a serial LAMMPS library build, which makes MLIP workflows impractical for my use case (I’d like to run evaluations in parallel).

Questions:

Is there any supported pathway to use DeePMD models directly or generally in eOn with/without going through LAMMPS?

Any pointers to the recommended workflow (or an example repository/script) would be very helpful.

Best regards,
Mohammed

[iz1ng]

Hi Mohammed, Yes, you can run deepmd through lammps, or as a python calculator through ase, or as an external potential. In all cases, I suggest having deepmd run on a gpu. So lammps can be serial, but deepmd should be built to run on a gpu. The same goes for ase and using an external potential. But unless you are trying to do an mpi cpu deepmd calculation, I don’t see any reason to have a parallel eon client. Best, Graeme

…

On Feb 25, 2026, at 7:57 AM, Mohammed Hassan ***@***.***> wrote: Hi eOn developers, I’m trying to use a DeePMD machine-learned interatomic potential (DeePMD-kit) with eOn for aKMC/rare-event searches. Context: I installed eOn 2.11.0 from conda-forge. My DeePMD potential is available as either: TensorFlow frozen model: model.pb, or PyTorch frozen model: model.pth (DeePMD v3 backend). I understand eOn can use LAMMPS as a potential interface, but my understanding is that eOn’s LAMMPS interface is typically set up with a serial LAMMPS library build, which makes MLIP workflows impractical for my use case (I’d like to run evaluations in parallel). Questions: Is there any supported pathway to use DeePMD models directly or generally in eOn with/without going through LAMMPS? Any pointers to the recommended workflow (or an example repository/script) would be very helpful. Best regards, Mohammed — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you are subscribed to this thread. This message is from an external sender. Learn more about why this matters.

[Mem314]

Regarding the three possibilities (LAMMPS, Python calculator through ASE, external potential), I now understand that using LAMMPS would require building eOn from source against a serial LAMMPS library. I would like to avoid that for now, since I want to stick with the conda-forge version of eOn.

For the other two options, which one should work with the conda build?

I tested the external potential option and observed the following:

eonclient works with potential = ext for job = point

python -m eon.server does not accept ext (it only accepts ext_pot)

When I switch to potential = ext_pot, the server starts, but the spawned client job fails with:
No known potential could be constructed from UNKNOWN
(as if eonclient does not recognize ext_pot)

I also tried potential = python, but when I run python -m eon.server I get:
No known potential could be constructed from PYTHON.

Do you have any suggestions for what I should try next? I’ve attached my scripts in case you’d like to take a closer look.

Thanks in advance.
Best, Mohammed

external_potential.py
config.txt

[HaoZeke]

This was a bug, now in v2.11.1 ext_pot should work correctly.

[HaoZeke]

There's better documentation for the ext_pot stuff now too: https://eondocs.org/user_guide/ext_pot

[Mem314]

Many thanks for the update. I tried it immediately and it worked well. As you mention in the documentation, however, the main limitation is the speed of each process, which is not ideal for my use case. Because of that, I decided to use the LAMMPS potential interface instead, which requires building from source.

I managed to build eOn with the LAMMPS interface, but I’m running into issues with DeepMD. In my current setup, LAMMPS does not recognize pair_style deepmd (I think). I tried using DeepMD in plugin mode, but I couldn’t get it to load cleanly.

Here is my in.lammps:
plugin load /home/moe/anaconda3/envs/eon/lib/python3.10/site-packages/deepmd/lib/libdeepmd_lmpplugin.so
pair_style deepmd /home/moe/eoN_lammps/models/frozen_model.pb
pair_coeff * * Al O

So i receive this error when running python -m eo.akmc or python -m eon.server:
----------------------------------------------------------------------------------------------------------#
Eon version 2.11.1
Registering results
Client running in .//jobs/in/0_0 returned no results. Check its output for errors.
Processed 0 results
Discarded approximately 1 results
Queue contains 0 searches
Making 1 process searches
job failed: .//jobs/scratch/0_1 (return code -6)
stderr: terminate called after throwing an instance of 'std::logic_error'
what(): basic_string: construction from null is not valid

stdout (last lines): * [Dimer] converged_angle: 5.0

• [Dimer] opt_method: cg
• [Dimer] rotations_max: 10
• [Prefactor] default_value: 1e12
• [Saddle Search] displace_magnitude: 0.15
• [Saddle Search] displace_radius: 12.0
• [Saddle Search] max_energy: 10.0
  Minimizing initial structure

[Matter] Iter Step size ||Force|| Energy
Created 1 searches
Currently in state 0 with confidence 0.000000
----------------------------------------------------------------------------------------------------------#

Here is my exact setup:

conda install -c conda-forge cmake make ninja meson compilers pkg-config

git clone https://github.com/lammps/lammps.git
cd lammps
mkdir -p build && cd build

cmake ../cmake
-D CMAKE_INSTALL_PREFIX="$CONDA_PREFIX"
-D BUILD_MPI=no
-D BUILD_SHARED_LIBS=yes
-D PKG_PLUGIN=yes
make -j
make install

conda install -c conda-forge eigen spdlog

rm -rf bbdir
meson setup bbdir -Dwith_lammps=True --prefix="$CONDA_PREFIX"

ninja -C bbdir
ninja -C bbdir install

[ERROR]: LAMMPSPot.cpp never includes LAMMPSPot.h, so i added manually the LAMMPSPot.h!

perl -0777 -i -pe 's/#include "LAMMPS.h"\s*\n#include "library.h"\s*\n/#include "LAMMPSPot.h"\n#include "LAMMPS.h"\n#include "lammps/library.h"\n#include \n#include \n/sg'
~/eOn/client/potentials/LAMMPS/LAMMPSPot.cpp
perl -0777 -i -pe 's/\n// LAMMPS library inclusion\s*\n#ifdef EONMPI\s*\n#define LAMMPS_LIB_MPI\s*\n#endif\s*\n#include "library.h"\s*\n/\n/sg'
~/eOn/client/potentials/LAMMPS/LAMMPSPot.cpp

ninja -C bbdir -t clean
ninja -C bbdir
ninja -C bbdir install

Thanks again for your help.

[Mem314]

Thank you for the quick response. That makes sense.

I’m trying to run aKMC with a DeePMD potential through the LAMMPS interface. My setup is:

config.ini
[Main]
job=akmc
temperature=300
checkpoint=true
finite_difference=0.1
random_seed=500

[Potential]
potential = lammps
log_potential = false
lammps_logging = true

[AKMC]
confidence=0.95
confidence_scheme=old
server_side_process_search=false
thermally_accessible_window=30

[Optimizer]
max_iterations = 200
opt_method = lbfgs
max_move = 0.05
converged_force = 0.01

[Structure Comparison]
distance_difference=0.2
energy_difference=0.1
neighbor_cutoff=3.4
check_rotation=true
indistinguishable_atoms=true

[Saddle Search]
displace_random_weight=1.0
displace_radius=12.0
displace_magnitude=0.15
max_energy=10.0

[Prefactor]
default_value=1e12

[Dimer]
opt_method=cg
converged_angle=5.0
rotations_max=10

[Debug]
keep_all_result_files=True
write_movies=True

pos.con
Generated by ASE
preBox_header_2
25.000000 25.000000 25.000000
90.000000 90.000000 90.000000
postBox_header_1
postBox_header_2
1
10
12.011
C
Coordinates of Component 1
106.186272 105.438910 106.773837 0 0
106.456366 104.599763 106.229101 0 1
105.470481 104.918846 106.226977 0 2
105.474695 105.955884 106.224370 0 3
106.459658 106.280396 106.227013 0 4
107.067777 105.441323 106.229104 0 5
105.909487 104.599761 105.344227 0 6
105.301468 105.438906 105.342188 0 7
105.912823 106.280394 105.342211 0 8
106.898766 105.961428 105.344250 0 9
106.896149 104.923121 105.345870 0 10

I also have a potfiles/ subdirectory containing in.lammps:

pair_style deepmd /home/eon/models/model.pth
pair_coeff * *

And my model is in models/ subdirectory containing model.pth

When I run python -m eon.server, I get:

Eon version 2.11.0
Registering results
Client running in .//jobs/in/0_0 returned no results. Check its output for errors.
Processed 0 results
Discarded approximately 1 results
Queue contains 0 searches
Making 1 process searches
job failed: .//jobs/scratch/0_1 (return code -6)
stderr: terminate called after throwing an instance of 'std::runtime_error'
what(): Terminating

stdout (last lines): ...
No known potential could be constructed from LAMMPS

As mentioned earlier, I installed eOn via conda:

conda install -c conda-forge eon

I would appreciate any guidance on how to proceed.

Best regards,
Mohammed

[HaoZeke]

Ah, right, you'll need to compile from source to get the LAMMPS interface, following these instructions:

https://eondocs.org/user_guide/lammps_pot

I'm also opening an issue, maybe the next conda release will link to LAMMPS (need to check how it works on Windows).

[HaoZeke]

For now the best is source compilation with LAMMPSPot and calling through LAMMPS, but for DeePMD specifically it should be possible to pass the model through one 0 training round of metatrain (from this PR) to generate a model (from the PyTorch frozen model) which can be loaded by Metatomic in eOn (the conda-forge one will work for this) [eondocs]

The code on conda-forge is ready for this and should be faster, the only problem is rendering the model in Metatomic format, which needs Metatrain at the moment.

The whole export and usage process is very much the same as the method to get MACE loaded by Metatomic.

This is not a very user-friendly / ergonomic workflow, I'll need to test it and can circle back maybe on Monday.