Getting Started

In this text, we will call the deep neural network that is used to represent the interatomic interactions (Deep Potential) the model. The typical procedure of using DeePMD-kit is

  1. Prepare data

  2. Train a model

  3. Freeze a model

  4. Test a model

  5. Compress a model

  6. Model inference

  7. Run MD

  8. Known limitations

Prepare data

One needs to provide the following information to train a model: the atom type, the simulation box, the atom coordinate, the atom force, system energy and virial. A snapshot of a system that contains these information is called a frame. We use the following convention of units:















The frames of the system are stored in two formats. A raw file is a plain text file with each information item written in one file and one frame written on one line. The default files that provide box, coordinate, force, energy and virial are box.raw, coord.raw, force.raw, energy.raw and virial.raw, respectively. We recommend you use these file names. Here is an example of force.raw:

$ cat force.raw
-0.724  2.039 -0.951  0.841 -0.464  0.363
 6.737  1.554 -5.587 -2.803  0.062  2.222
-1.968 -0.163  1.020 -0.225 -0.789  0.343

This force.raw contains 3 frames with each frame having the forces of 2 atoms, thus it has 3 lines and 6 columns. Each line provides all the 3 force components of 2 atoms in 1 frame. The first three numbers are the 3 force components of the first atom, while the second three numbers are the 3 force components of the second atom. The coordinate file coord.raw is organized similarly. In box.raw, the 9 components of the box vectors should be provided on each line. In virial.raw, the 9 components of the virial tensor should be provided on each line in the order XX XY XZ YX YY YZ ZX ZY ZZ. The number of lines of all raw files should be identical.

We assume that the atom types do not change in all frames. It is provided by type.raw, which has one line with the types of atoms written one by one. The atom types should be integers. For example the type.raw of a system that has 2 atoms with 0 and 1:

$ cat type.raw
0 1

Sometimes one needs to map the integer types to atom name. The mapping can be given by the file type_map.raw. For example

$ cat type_map.raw

The type 0 is named by "O" and the type 1 is named by "H".

The second format is the data sets of numpy binary data that are directly used by the training program. User can use the script $deepmd_source_dir/data/raw/ to convert the prepared raw files to data sets. For example, if we have a raw file that contains 6000 frames,

$ ls 
box.raw  coord.raw  energy.raw  force.raw  type.raw  virial.raw
$ $deepmd_source_dir/data/raw/ 2000
nframe is 6000
nline per set is 2000
will make 3 sets
making set 0 ...
making set 1 ...
making set 2 ...
$ ls 
box.raw  coord.raw  energy.raw  force.raw  set.000  set.001  set.002  type.raw  virial.raw

It generates three sets set.000, set.001 and set.002, with each set contains 2000 frames. One do not need to take care of the binary data files in each of the set.* directories. The path containing set.* and type.raw is called a system.

Data preparation with dpdata

One can use the a convenient tool dpdata to convert data directly from the output of first priciple packages to the DeePMD-kit format. One may follow the example of using dpdata to find out how to use it.

Train a model

Write the input script

A model has two parts, a descriptor that maps atomic configuration to a set of symmetry invariant features, and a fitting net that takes descriptor as input and predicts the atomic contribution to the target physical property.

DeePMD-kit implements the following descriptors:

  1. se_e2_a: DeepPot-SE constructed from all information (both angular and radial) of atomic configurations. The embedding takes the distance between atoms as input.

  2. se_e2_r: DeepPot-SE constructed from radial information of atomic configurations. The embedding takes the distance between atoms as input.

  3. se_e3: DeepPot-SE constructed from all information (both angular and radial) of atomic configurations. The embedding takes angles between two neighboring atoms as input.

  4. loc_frame: Defines a local frame at each atom, and the compute the descriptor as local coordinates under this frame.

  5. hybrid: Concate a list of descriptors to form a new descriptor.

The fitting of the following physical properties are supported

  1. ener: Fitting the energy of the system. The force (derivative with atom positions) and the virial (derivative with the box tensor) can also be trained. See the example.

  2. dipole: The dipole moment.

  3. polar: The polarizability.


The training can be invoked by

$ dp train input.json

where input.json is the name of the input script. See the example for more details.

During the training, checkpoints will be written to files with prefix save_ckpt every save_freq training steps.

Several command line options can be passed to dp train, which can be checked with

$ dp train --help

An explanation will be provided

positional arguments:
  INPUT                 the input json database

optional arguments:
  -h, --help            show this help message and exit
  --init-model INIT_MODEL
                        Initialize a model by the provided checkpoint
  --restart RESTART     Restart the training from the provided checkpoint

--init-model model.ckpt, initializes the model training with an existing model that is stored in the checkpoint model.ckpt, the network architectures should match.

--restart model.ckpt, continues the training from the checkpoint model.ckpt.

On some resources limited machines, one may want to control the number of threads used by DeePMD-kit. This is achieved by three environmental variables: OMP_NUM_THREADS, TF_INTRA_OP_PARALLELISM_THREADS and TF_INTER_OP_PARALLELISM_THREADS. OMP_NUM_THREADS controls the multithreading of DeePMD-kit implemented operations. TF_INTRA_OP_PARALLELISM_THREADS and TF_INTER_OP_PARALLELISM_THREADS controls intra_op_parallelism_threads and inter_op_parallelism_threads, which are Tensorflow configurations for multithreading. An explanation is found here.

For example if you wish to use 3 cores of 2 CPUs on one node, you may set the environmental variables and run DeePMD-kit as follows:

dp train input.json

Training analysis with Tensorboard

If enbled in json/yaml input file DeePMD-kit will create log files which can be used to analyze training procedure with Tensorboard. For a short tutorial please read this document.

Freeze a model

The trained neural network is extracted from a checkpoint and dumped into a database. This process is called “freezing” a model. The idea and part of our code are from Morgan. To freeze a model, typically one does

$ dp freeze -o graph.pb

in the folder where the model is trained. The output database is called graph.pb.

Test a model

The frozen model can be used in many ways. The most straightforward test can be performed using dp test. A typical usage of dp test is

dp test -m graph.pb -s /path/to/system -n 30

where -m gives the tested model, -s the path to the tested system and -n the number of tested frames. Several other command line options can be passed to dp test, which can be checked with

$ dp test --help

An explanation will be provided

usage: dp test [-h] [-m MODEL] [-s SYSTEM] [-S SET_PREFIX] [-n NUMB_TEST]
               [-r RAND_SEED] [--shuffle-test] [-d DETAIL_FILE]

optional arguments:
  -h, --help            show this help message and exit
  -m MODEL, --model MODEL
                        Frozen model file to import
  -s SYSTEM, --system SYSTEM
                        The system dir
  -S SET_PREFIX, --set-prefix SET_PREFIX
                        The set prefix
  -n NUMB_TEST, --numb-test NUMB_TEST
                        The number of data for test
  -r RAND_SEED, --rand-seed RAND_SEED
                        The random seed
  --shuffle-test        Shuffle test data
  -d DETAIL_FILE, --detail-file DETAIL_FILE
                        The file containing details of energy force and virial

Compress a model

Once the frozen model is obtained from deepmd-kit, we can get the neural network structure and its parameters (weights, biases, etc.) from the trained model, and compress it in the following way:

dp compress input.json -i graph.pb -o graph-compress.pb

where input.json denotes the original training input script, -i gives the original frozen model, -o gives the compressed model. Several other command line options can be passed to dp compress, which can be checked with

$ dp compress --help

An explanation will be provided

usage: dp compress [-h] [-i INPUT] [-o OUTPUT] [-e EXTRAPOLATE] [-s STRIDE]
                   [-f FREQUENCY] [-d FOLDER]

positional arguments:
  INPUT                 The input parameter file in json or yaml format, which
                        should be consistent with the original model parameter

optional arguments:
  -h, --help            show this help message and exit
  -i INPUT, --input INPUT
                        The original frozen model, which will be compressed by
                        the deepmd-kit
  -o OUTPUT, --output OUTPUT
                        The compressed model
                        The scale of model extrapolation
  -s STRIDE, --stride STRIDE
                        The uniform stride of tabulation's first table, the
                        second table will use 10 * stride as it's uniform
  -f FREQUENCY, --frequency FREQUENCY
                        The frequency of tabulation overflow check(If the
                        input environment matrix overflow the first or second
                        table range). By default do not check the overflow
  -d FOLDER, --folder FOLDER
                        path to checkpoint folder

Parameter explanation

Model compression, which including tabulating the embedding-net. The table is composed of fifth-order polynomial coefficients and is assembled from two sub-tables. The first sub-table takes the stride(parameter) as it’s uniform stride, while the second sub-table takes 10 * stride as it’s uniform stride. The range of the first table is automatically detected by deepmd-kit, while the second table ranges from the first table’s upper boundary(upper) to the extrapolate(parameter) * upper. Finally, we added a check frequency parameter. It indicates how often the program checks for overflow(if the input environment matrix overflow the first or second table range) during the MD inference.

Justification of model compression

Model compression, with little loss of accuracy, can greatly speed up MD inference time. According to different simulation systems and training parameters, the speedup can reach more than 10 times at both CPU and GPU devices. At the same time, model compression can greatly change the memory usage, reducing as much as 20 times under the same hardware conditions.

Acceptable original model version

The model compression method requires that the version of DeePMD-kit used in original model generation should be 1.3 or above. If one has a frozen 1.2 model, one can first use the convenient conversion interface of DeePMD-kit-v1.2.4 to get a 1.3 executable model.(eg: dp convert-to-1.3 -i frozen_1.2.pb -o frozen_1.3.pb)

Model inference

Python interface

One may use the python interface of DeePMD-kit for model inference, an example is given as follows

from deepmd.infer import DeepPot
import numpy as np
dp = DeepPot('graph.pb')
coord = np.array([[1,0,0], [0,0,1.5], [1,0,3]]).reshape([1, -1])
cell = np.diag(10 * np.ones(3)).reshape([1, -1])
atype = [1,0,1]
e, f, v = dp.eval(coord, cell, atype)

where e, f and v are predicted energy, force and virial of the system, respectively.

C++ interface

The C++ interface of DeePMD-kit is also avaiable for model interface, which is considered faster than Python interface. An example infer_water.cpp is given below:

#include "deepmd/DeepPot.h"

int main(){
  deepmd::DeepPot dp ("graph.pb");
  std::vector<double > coord = {1., 0., 0., 0., 0., 1.5, 1. ,0. ,3.};
  std::vector<double > cell = {10., 0., 0., 0., 10., 0., 0., 0., 10.};
  std::vector<int > atype = {1, 0, 1};
  double e;
  std::vector<double > f, v;
  dp.compute (e, f, v, coord, atype, cell);

where e, f and v are predicted energy, force and virial of the system, respectively.

You can compile infer_water.cpp using gcc:

gcc infer_water.cpp -D HIGH_PREC -L $deepmd_root/lib -L $tensorflow_root/lib -I $deepmd_root/include -I $tensorflow_root/lib -Wl,--no-as-needed -ldeepmd_op -ldeepmd -ldeepmd_cc -ltensorflow_cc -ltensorflow_framework -lstdc++ -Wl,-rpath=$deepmd_root/lib -Wl,-rpath=$tensorflow_root/lib -o infer_water

and then run the program:


Run MD

Run MD with LAMMPS

Include deepmd in the pair_style


pair_style deepmd models ... keyword value ...
  • deepmd = style of this pair_style

  • models = frozen model(s) to compute the interaction. If multiple models are provided, then the model deviation will be computed

  • keyword = out_file or out_freq or fparam or atomic or relative

    out_file value = filename
        filename = The file name for the model deviation output. Default is model_devi.out
    out_freq value = freq
        freq = Frequency for the model deviation output. Default is 100.
    fparam value = parameters
        parameters = one or more frame parameters required for model evaluation.
    atomic = no value is required. 
        If this keyword is set, the model deviation of each atom will be output.
    relative value = level
        level = The level parameter for computing the relative model deviation


pair_style deepmd graph.pb
pair_style deepmd graph.pb fparam 1.2
pair_style deepmd graph_0.pb graph_1.pb graph_2.pb out_file md.out out_freq 10 atomic relative 1.0


Evaluate the interaction of the system by using Deep Potential or Deep Potential Smooth Edition. It is noticed that deep potential is not a “pairwise” interaction, but a multi-body interaction.

This pair style takes the deep potential defined in a model file that usually has the .pb extension. The model can be trained and frozen by package DeePMD-kit.

The model deviation evalulate the consistency of the force predictions from multiple models. By default, only the maximal, minimal and averge model deviations are output. If the key atomic is set, then the model deviation of force prediction of each atom will be output.

By default, the model deviation is output in absolute value. If the keyword relative is set, then the relative model deviation will be output. The relative model deviation of the force on atom i is defined by

Ef_i = -------------
       |f_i| + level

where Df_i is the absolute model deviation of the force on atom i, |f_i| is the norm of the the force and level is provided as the parameter of the keyword relative.


  • The deepmd pair style is provided in the USER-DEEPMD package, which is compiled from the DeePMD-kit, visit the DeePMD-kit website for more information.

Long-range interaction

The reciprocal space part of the long-range interaction can be calculated by LAMMPS command kspace_style. To use it with DeePMD-kit, one writes

pair_style	deepmd graph.pb
kspace_style	pppm 1.0e-5
kspace_modify	gewald 0.45

Please notice that the DeePMD does nothing to the direct space part of the electrostatic interaction, because this part is assumed to be fitted in the DeePMD model (the direct space cut-off is thus the cut-off of the DeePMD model). The splitting parameter gewald is modified by the kspace_modify command.

Run path-integral MD with i-PI

The i-PI works in a client-server model. The i-PI provides the server for integrating the replica positions of atoms, while the DeePMD-kit provides a client named dp_ipi that computes the interactions (including energy, force and virial). The server and client communicates via the Unix domain socket or the Internet socket. Installation instructions of i-PI can be found here. The client can be started by

i-pi input.xml &
dp_ipi water.json

It is noted that multiple instances of the client is allow for computing, in parallel, the interactions of multiple replica of the path-integral MD.

water.json is the parameter file for the client dp_ipi, and an example is provided:

    "verbose":		false,
    "use_unix":		true,
    "port":		31415,
    "host":		"localhost",
    "graph_file":	"graph.pb",
    "coord_file":	"",
    "atom_type" : {
	"OW":		0, 
	"HW1":		1,
	"HW2":		1

The option use_unix is set to true to activate the Unix domain socket, otherwise, the Internet socket is used.

The option port should be the same as that in input.xml:


The option graph_file provides the file name of the frozen model.

The dp_ipi gets the atom names from an XYZ file provided by coord_file (meanwhile ignores all coordinates in it), and translates the names to atom types by rules provided by atom_type.

Use deep potential with ASE

Deep potential can be set up as a calculator with ASE to obtain potential energies and forces.

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],

Optimization is also available:

from ase.optimize import BFGS
dyn = BFGS(water)

Known limitations

If you use deepmd-kit in a GPU environment, the acceptable value range of some variables are additionally restricted compared to the CPU environment due to the software’s GPU implementations:

  1. The number of atom type of a given system must be less than 128.

  2. The maximum distance between an atom and it’s neighbors must be less than 128. It can be controlled by setting the rcut value of training parameters.

  3. Theoretically, the maximum number of atoms that a single GPU can accept is about 10,000,000. However, this value is actually limited by the GPU memory size currently, usually within 1000,000 atoms even at the model compression mode.

  4. The total sel value of training parameters(in model/descriptor section) must be less than 4096.