Computing energy

[lukanov-9b]

Hello, dear Zheyong Fan!

Could you tell me please, how can I compute energies of structures separately? I created a model.xyz-file with structure configurations (each structure about 64 atoms). After that I used nep-potential (nep.txt) file for computing, indicating prediction mode as 1. But I have some results:
(First column - prediction, second column - real values)
-12.2512 -21.6235
-14.1193 -21.2266
-14.2669 -20.5895
-14.0811 -20.6964
-12.0444 -20.6308
-15.2373 -19.5306
-13.6627 -19.9599
-11.9918 -20.9433
-13.6622 -19.9643
-13.6825 -19.9623

Structures were taken from test.xyz file used for training process. I can't understand, why errors in this case, when I predict energies separately from training, are significant. Thanks!

[brucefan1983]

It's not very clear from your description what you did exactly. I guess you mean you used the prediction 1 keyword in the nep.in file to predict results for a train.xyz file, where the structures in train.xyz are some new structures you want to predict?

[brucefan1983]

If so, you can post figures showing the training and testing results. Also the nep.in file. The more information the better.

[lukanov-9b]

So, I have some structures and I need to know energies of them. I read that I can use nep-code for this purpose if I have nep.txt file in the directory and prediction 1 key-word. I wanted to check this on structures that were used ifor test.xyz-file, when I trained nep-potential. And I have different errors...

[lukanov-9b]

When I trained model

If I do this from prediction 1, aI have

[brucefan1983]

Could you please post the two nep.in files, one for training, one for testing (with prediciton 1).

Also, could you tell me the versions of GPUMD if you did them using different versions.

[lukanov-9b]

The first nep.in file:
type 5 Hf Nb Ta Ti Zr

The second nep.in file:
type 5 Hf Nb Ta Ti Zr
prediction 1

[lukanov-9b]

I have Version 3.8

[brucefan1983]

Then you can send me the inputs to debug (via email).

[brucefan1983]

You must have used different versions of GPUMD for training and prediction and the two versions happen to have different default values for basis_size.

If you add basis_size 12 12 to your nep.in and do the predition again, you will get the correct results.

[brucefan1983]

I got basis_size 12 12 from your nep.txt. The point is that nep.in (with many default parameters that might be changed from version to version) should be consistent with nep.txt to give correct predictions.

[lukanov-9b]

Hello, dear Zheyong Fan! It has worked, thank you so much!!!

I wanted ask you about one moment of MC. Could you tell me, please, what difference are between
<run ...> and <mc_trials>

For example, I use this configuration:
mc canonical 1 1000 500 500
time_step 0
dump_exyz 1000
run 1000000

GPUMD does 1000 steps of MC in each step (from 1000000 in run-option)? When I used mc canonical 1 1 500 500 I got 1 or 0 in mcmd.out. But when I use more then 1 mc trials, I have float values in mcmd.out.

[brucefan1983]

Hello, I think you can understand everything by reading the following two pages:

• https://gpumd.org/gpumd/input_parameters/mc.html

• https://gpumd.org/gpumd/output_files/mcmd_out.html#mcmd-out