Exception: No weights file found for element: O

[ZengZezhu]

Dear AML Team,

I just want to use the aml to generate a NNP potential of water. When I used the script of "create-model.py" prepared in the examples, I encountered an issue like below:

Traceback (most recent call last):
File "/lustre1/u/u3004964/using_n2p2/aml-committee/water-cheng/create-model.py", line 35, in
n2p2.train(structures, n_epoch=2)
File "/lustre1/u/u3004964/software/aml/aml/mlp.py", line 206, in train
self._read_parameters(directories)
File "/lustre1/u/u3004964/software/aml/aml/mlp.py", line 160, in _read_parameters
model_parameters.append(self._read_output_train(directory))
File "/lustre1/u/u3004964/software/aml/aml/mlp.py", line 449, in _read_output_train
raise Exception(f'No weights file found for element: {element:s}')
Exception: No weights file found for element: O

I carefully prepared the input files (including input.data, input.nn and scaling,data) for training the NNP by using n2p2, and I also tested and found that it is correct if I directly run 'nnp-train' in the path of 'final-training/train-000'. Could you tell me what should I do for this issue? Many thanks.

I am also confused about the aml function, does aml can be sued to train the NNP potential? or we need to use the n2p2 to train and then use the aml to generate the C-NNP?

Cheers,
zezhu

[ZengZezhu]

Hi,

Now I solved this issue by a pre-fitting when performing the create-model.py, so I think that aml cannot be sued to fit the NNP.

But I am still a little confused by the MD simulation of using C-NNP in lammps. I see that the final model just is a simple combination of 8 NNPs located on 8 different folders. Is it possible to use the final C-NNP model to perform MD simulations in lammps? Because I do not find the final C-NNP potential in the model folder, I don't know what should I do to average these 8 NNPs. I see the example of MD simulations in '04-C-NNP-MD' only gives the input of CP2K.:)

[cschran]

Committee NNPs are currently only available in CP2K, but there is an open n2p2 PR which implements committees for lammps. This PR can already be used.

[cschran]

Hi Zezhu,

thanks for your interest in our work.

AML can be faithfully used to train a model with n2p2. This is a crucial component of our active learning process.

Internally, AML calls nnp-scaling/train, but you need to make sure that both are in your path. Furthermore, mpirun might cause problems in your particular case. To trouble shoot, try running each training process on a single core (n_core_task = 1).

[OndrejMarsalek]

Just to further explain, the current version of the code is quite basic in that it does not do proper error checking on the subprocesses (like n2p2 executables) that it launches. That means that if they fail for whatever reason, you only get an error once AML tries to read their output and does not find it. As Christoph says, the most common reason for this failure is the runtime environment - the executable is not found, MPI does not work properly, or something similar. You want to make sure you can run n2p2 "by hand" in the same environment where you're running AML.

The development version of AML that we should be making public soon fixes this and will have better error reporting when subprocesses fail.

[ZengZezhu]

Hi,

Thank you for your information! I got it.

Cheers,
zezhu