[BUG] Inconsistent optimization trajectories when using DPA2 and ASE

[Chengqian-Zhang]

Bug summary

When using DPA2 and ASE for two identical structural optimizations, the trajectories are inconsistent.

The first time:

The second time:

Note that the python script and init structure is completely the same. So the two trajs should be the same.

DeePMD-kit Version

devel

Backend and its version

pytorch v2.5.1+cu124

How did you download the software?

pip

Input Files, Running Commands, Error Log, etc.

# relax.py
from ase import Atoms
from deepmd.calculator import DP
from ase.io import Trajectory
water = Atoms('H2O3', positions=[(0.3, 0.3, 0.3),(0.5, 0.5, 0.5),(0.7, 0.7, 0.7),(0.1,0.1,0.1),(0.6,2.5,3.7)],cell=[5, 5, 5],calculator=DP(model="dpa2_frozen_model.pth"))
print(water.get_potential_energy())
print(water.get_forces())
from ase.optimize import LBFGS
traj = Trajectory('traj.traj', 'w', water)
dyn = LBFGS(water)
dyn.attach(traj.write, interval=1)
dyn.run(fmax=1e-6,steps=50)

print(water.get_positions())

Steps to Reproduce

run

python3 relax.py

for two times.
The results are different.

Further Information, Files, and Links

When a structure is used for energy and force inference, there may be some small errors (such as 1e-10), but this error of dpa2 will accumulate with the increase of optimization steps, and this bug will occur when it accumulates to a relatively large value. The error accumulation of the small model (se_2_a) is not as fast as dpa2.

[njzjz]

Could you call torch.use_deterministic_algorithms(True) and see if the result is still different?

[wanghan-iapcm]

It may happen that starting from the same initial point, the structure is relaxed to different local minima due to the small perturbation (like the truncation error) on the trajectory.

water = Atoms('H2O3', positions=[(0.3, 0.3, 0.3),(0.5, 0.5, 0.5),(0.7, 0.7, 0.7),(0.1,0.1,0.1),(0.6,2.5,3.7)],cell=[5, 5, 5],calculator=DP(model="dpa2_frozen_model.pth"))

It is not physically meaningful structure,

1. the formula does not make sense.
2. at least the first 2 H atoms are too close: distance 0.34 A

[QuantumMisaka]

@Chengqian-Zhang You can try with a physical input structure. in my practices using DPA-2 for surface reactions simulation, the geometry optimizaition and even transition state optimization in ASE using DPA-2 will be repeatable

[Chengqian-Zhang]

Thank you for your patient explanation.
There are two conclusion:

1. Reaching different local minima due to the small perturbation is normal whening using unphysically initial structure.
2. If one want to reach completely the same local minima. Calling torch.use_deterministic_algorithms(True) works well.
   So I will close this issue.