Building a Moment Tensor Potential

Building an MTP (Moment Tensor Potential)

Constructing a Moment Tensor potential is an iterative process wherein the model is repeatedly refined and improved until it predicts well across the entire search space. To make this iterative process easier, a python package called aBuild(see here for installation instructions) has been built. aBuild automates file building and information extraction that is constantly being done during the model-building process. Below you will find the steps needed to build an MTP, using aBuild as the automation tool. Anytime you run aBuild (through the builder.py script) you will need a yaml file to specify run settings. Here is an example yaml file. Make sure you modify it for your system. Note: In the instructions below, the name of my yaml file was AgAu.yml. You should modify the command for your system.

1. Enumerate derivative superstructures

In order to build the library of candidate ground states, we need to enumerate the candidates. See the derivative superstructures section here for more information. The command to perform the enumeration is:

python builder.py AgAu -enum -dataset='gss'

(Note: the “dataset” option to the command specifiies which dataset in the YAML file will be used to perform the enumeration) When this runs it will pull information from the ‘gss’ section of your yaml file to determine which parent lattices to enumerate on. After this completes, you should have a folder called Enum with the enumeration files in there. (struct_enum.out.<lattice>) You should only have to do this step once per system.

2. Define the search space

Now it’s time to build a file that contains all of the crystal structures that are considered to be candidate ground states. It’s important that this file be as exhaustive as possible, which means we should include any crystal structure that we suspect might be low in energy.(Likely, this means hundreds of thousands or even millions of crystal structures!) The code to generate this file is:

python builder.py AgAu -setup_relax

When this runs, it will use the information in the ‘gss’ section of the yaml file to generate the crystal structures. Due to the large number of crystals that will be generated, this command may take several hours to complete. You can either i) submit a job and let it run overnight or ii) split up the file generation over many processors and let it finish in less than half an hour. (Here is an example submission script that will do that for you) When this finishes, you will have a folder called fitting/mtp. Inside that folder you will notice the following files:

to_relax.cfg
relax.ini
pot.mtp
train.cfg
jobscript_relax.sh

(If you split up the job over many processors, you’ll have to concatenate all of the files into one, like this:

cd fitting/mtp
cat to_relax.cfg_* > to_relax.cfg
cd -

)

3. Relax

Once step 2 finishes, there will be a job script located in fitting/mtp that you can submit. This job will use the MTP (we haven’t generated one yet, but that’s OK) to attempt to relax the atoms in all of the crystals.

cd fitting/mtp
sbatch jobscript_relax.sh

4. Select

As the model relaxes the atoms, it will find that it has to extrapolate on some of them. In other words, the model is making calculations on a crystal structure that is very different from the ones it was trained on. It saves these structures and selects a subset of them to be added to the training set:

python builder.py AgAu -setup_select_add

This will only take a few seconds to run and when it’s finished there will be a new job submission file located in fitting/mtp called: jobscript_select.sh. Submit it:

cd fitting/mtp
sbatch jobscript_select.sh

and wait for it to complete. This could take several hours to complete and, in later iterations, this job may need the better part of a day to complete.

5. Augment Training Set

Now we need to build VASP folders for the structures that were requested be added into the training set:

python builder.py AgAu -add

This will generate several hundred folders inside of the folder training_set. Each folder has the needed files to perform a quantum-mechanical calculation for a different crystal. There will also be a job submission file located in training_set called jobscript_vasp.sh. Submit the job:

cd training_set
sbatch jobscript_vasp.sh

and let the calculations run. Each job will need anywhere from a couple of hours to days to run.

Calculations that have errors.

Sometimes a calculation will not finish correctly for one reason or another. aBuild has a mode called -status that will check all of the calculation folders for errors. You can run it like this:

python3 builder.py AgAu -status

This will create files entitled status.<runcode>, where <runcode> can be one of : running, done, unconverged, and errors. If you have unconverged or erroneous calculations, you must investigate the problem and restart the calculation before moving on. See [this page][vaspdebug] for help finding and fixing vasp errors.

6. Build Training Set

Once the calculations are finished, we need to extract the results and compile the data into the training set file so we can train the model:

python builder.py AgAu -setup_train

This will generate the following files in fitting/mtp:

train.cfg
relax.ini
state.mvs
jobscript_train.sh

Submit the job submission script:

cd fitting/mtp
sbatch jobscript_train.sh

and wait for it to finish. Once finished you will notice the file:

Trained.mtp_

located in fitting/mtp. This file contains values for the model parameters.

7. Return to step #2

With a trained potential, we can return to step #2 and attempt to relax again. Look from step #2 to step #7 until the model successfully relaxes all of the crystals in to_relax.cfg.