Compute pe/atom in LAMMPS interface does not include normalization offset

[owenrett]

Describe the bug The command compute pe/atom (atom specific energy) in lammps returns the potential energy without either the normalization offset ("mean_energy"), nor the atom energy offset ("atom_energy"). The command compute pe (total system energy) appears to include these offsets. For a small scale system (96 atoms, 3 species, Zr Ce O) the energies returned by compute pe/atom (reduced and summed) are:

Relevant input.nn Settings mean_energy: -3.3958016388641306E+00 atom_energy Zr -7.73537698 atom_energy Ce -4.63794868 atom_energy O -1.54279418

System Setup N_Zr N_Ce N_O 28 4 64

LAMMPS: computing energies per atom compute ZrEnA ZrGp pe/atom compute ZrEn ZrGp reduce sum c_ZrEnA

ZrEn = -210.00812 CeEn = 4.1838487 OEn = 211.13337

Computing total energy compute peAll all pe -654.24414

Comparing Zr_En_Sum + N_Zr Mean_Energy + N_Zr atom_energy_Zr = -210.00812 + 28 -3.395 + 28 -7.735 = -522.715 Ce_En_Sum + N_Ce Mean_Energy + N_Ce atom_energy_Ce = 4.1838487 + 4 -3.395 + 4 -4.637 = -27.584 O_En_Sum + N_O Mean_Energy + N_O atom_energy_O = 211.13337 + 64 -3.395 + 64 - 1.543 = -103.944

(-522.715) + (-27.584) + (-103.944) = -654.24414

To Reproduce Steps to reproduce the behavior:

1. Which git commit version is used? commit 293a6a2
2. What are the compilation options (compiler, flags)? Standard Intel Compilation with: PROJECT_CFLAGS=-O3 -xHost -std=c++11 -ipo -xSSSE3 -axSSE4.2,AVX,CORE-AVX-I,CORE-AVX2
3. Which application is affected? LAMMPS-n2p2 interface
4. What are the actual settings (please provide a minimal example if possible)? I have attached the input.nn, md.lmp, and log.lammps files.
5. Please provide the error message or describe the crash behavior. No Error Message given; compute pe/atom does not match compute pe when summed

Expected behavior A clear and concise description of what you expected to happen. Expect compute pe/atom to match `compute pe when per atom energies are summed. This issue can be fixed in post if energies calculated are necessary, however I figured that it should be brought to your attention.

Additional context The relevant files have been attached below.

md_NETO.lmp.txt log.lammps.txt input.nn.txt

---

My apologies if the information here is not enough or formatted incorrectly. I have not submitted many issues through github before. I had drafted this issue for a significantly larger system before reporting for a smaller system, so some numbers may be slightly off; I am trying to fix them.

-Owen Rettenmaier

[owenrett]

I realize I forgot to attach the lammps structure file to the issue. I have attached it here.

Ce12_M_S.lmp.txt

[singraber]

Hello Owen!

thank you very much for this carefully prepared bug report, your effort makes it really much easier for me to follow the description and trace back the origin. I greatly appreciate that :-)!

I just submitted a commit (https://github.com/CompPhysVienna/n2p2/commit/e3ee6abef466d8d313f5b1d4bda421a444d3e4a0) to the master branch which should fix the problem, could you please test if that is the case? (I cannot because I would need also the weights.???.data and scaling.data files.)

Thanks a lot for catching this!

Best, Andreas Singraber

[owenrett]

My apologies on the delay; had a bug on my system related to openmpi, so took a while to recompile. The energies on a per atom and system basis seem to agree now; that is both are giving equivalent values.