Skip to content

Latest commit

 

History

History
135 lines (95 loc) · 7.03 KB

README.md

File metadata and controls

135 lines (95 loc) · 7.03 KB
Raw

DeePMD

1. Introduction

Example(s) in this folder demonstrate how to use DeePMD (Deep Potential Molecular Dynamics)1, a package for building neural network potentials for molecular dynamics. The trained models can be interfaced directly to LAMMPS and other molecular dynamics codes.

1.1. Example(s): Machine Learning Force-field for H2O.

Located in train_mlff_qe-cp-traj. In this CLI job example, we perform:

  1. Ab-initio molecular dynamics calculation using Quantum ESPRESSO cp.x
  2. Transform output data into DeePMD format for training and validation
  3. Train a model and freeze it (save into a file) for reuse.
  4. Use the frozen model for molecular dynamics calculation using LAMMPS included with DeePMD distribution.

The setup is similar to the one in Ref. 2 below. The main difference is that we use the cp.x program from Quantum ESPRESSO to perform ab-initio molecular dynamics calculation. The resulting trajectory is then used to train a model using DeePMD. The trained model is then used to perform molecular dynamics calculation using LAMMPS.

The workflow can be invoked using the run.sh script in the top-level folder. The job.pbs file in the top-level folder can be submitted to the queue to run all steps in sequence using run.sh.

Example usage (for a user inside the Mat3ra.com CLI):

mkdir mlff-h2o && cd mlff-h2o
cp -r ~/job_script_templates/deepmd/train_mlff_qe-cp-traj .
cd train_mlff_qe-cp-traj
sh run.sh

Then, we can check the status of the job using qstat command. Read more about the command-line jobs in the documentation. Note that although 4 jobs are submitted, only one job is running at a time. The other jobs are in the queue waiting for the previous job in the sequence to finish.

qstat
Job ID                    Name             User            Time Use S Queue
------------------------- ---------------- --------------- -------- - -----
94883.master-production-20160 ...-MLFF-step_01 timur           00:00:01 R D              
94884.master-production-20160 ...-MLFF-step_02 timur                  0 H D              
94885.master-production-20160 ...-MLFF-step_03 timur                  0 H D              
94886.master-production-20160 ...-MLFF-step_04 timur                  0 H D

The output directories can then be monitored with

watch -n 5 ls -tlhra */output/

Note

To have examples that can be run rather quickly, the setup is such that the model in question is trained with a small amount of data for a short period of time. Consequently, the accuracy of the resulting model is low as intended.

2. "DeePMD-MLFF" Workflow Steps

Important notes:

  • We isolate each step to the corresponding folder starting with step_ prefix, e.g. `step_01_generate_dft_data. This way we can run each step independently.
  • Each subsequent step depends on the previous step. For example, step_02_preprocess_dft_data depends on step_01_generate_dft_data and so on. The content of the output folder of the previous step is used as the input for the next step by linking the files. See ./step_02_preprocess_dft_data/job.pbs for an example.
  • We use the Folder Structure with "input", "output", and "output-reference" folders for each step.
  • Each step has a job.pbs file that can be submitted to the queue independently, assuming the previous step has been completed.
  • There is a top-level run.sh script that runs all steps in sequence. The job.pbs file in the top-level folder can be submitted to the queue to run all steps in sequence using run.sh.
  • In order to introduce modifications, one can also edit settings.sh file in the top-level folder.

2.1. Ab-initio Molecular Dynamics.

Perform ab-initio molecular dynamics calculation using Quantum ESPRESSO cp.x program. For input files see under espresso/cp.x. We can run the calculation by using Car-Parrinello molecular dynamics with Quantum Espresso similar to example from Ref 3.

Note

We have set prefix "generate_dft_data" to be the same as the base input filename "generate_dft_data.in". This way we get the same base file names ({prefix}.???) for other output files. When loading data using DeepMD dpdata tool, it expects the base names for various input files to be the same.

2.2. Data Conversion and Train/Test Split.

Next we invoke a python script to transform the output of a Car-Parrinello molecular dynamics run using Quantum Espresso into the input format used by DeePMD-Kit. The data are saved under model_training/data. We divide dataset into two parts: 80 percent of the frames for training and 20 percent of the frames for validation.

2.3. Model Training

The training script trains a model on the water data prepared in the previous step and then freezes that model for production. The job must be submitted from the same directory as the model_training/job.pbs file. The frozen model is saved to file model_training/results/graph.pb.

2.4. (Classical) Molecular Dynamics.

The molecular dynamics script takes the graph.pb file produced by the training example and carries out a prospective simulation using LAMMPS. As above the job must be submitted from the same directory as the molecular_dynamics/job.pbs file.

2.5. (Extra) Model inference

We can use the DeePMD python interface for model inference, for example, we can evaluate the energy of each frame in the LAMMPS trajectory.

import dpdata

data = dpdata.System('structure.dump', fmt="lammps/dump", type_map=["O", "H"])
d_predict = data.predict(dp="../model_training/results/graph.pb")
print(d_predict["energies"])

2.6. (Extra) Molecular Dynamics with ASE.

We can also use the DeePMD python interface for invoking the Atomistic Simulations Environment (ASE) as explained in 2).

from ase import Atoms
from deepmd.calculator import DP

water = Atoms('H2O',
              positions=[(0.7601, 1.9270, 1),
                         (1.9575, 1, 1),
                         (1., 1., 1.)],
              cell=[100, 100, 100],
              calculator=DP(model="frozen_model.pb"))
print(water.get_potential_energy())
print(water.get_forces())

Optimization can be performed as:

from ase.optimize import BFGS
dyn = BFGS(water)
dyn.run(fmax=1e-6)
print(water.get_positions())

3. Links.

Footnotes

  1. DeePMD-kit: A deep learning package for many-body potential energy representation and molecular dynamic; (arXiv). ↩

  2. DeePMD-kit: Getting Started ↩ ↩2

  3. Quantum ESPRESSO: cp.x User guide ↩