3.5. Descriptor "se_atten"

3.5.1. DPA-1: Pretraining of Attention-based Deep Potential Model for Molecular Simulation


Here we propose DPA-1, a Deep Potential model with a novel attention mechanism, which is highly effective for representing the conformation and chemical spaces of atomic systems and learning the PES.

See this paper for more information. DPA-1 is implemented as a new descriptor "se_atten" for model training, which can be used after simply editing the input.json.

3.5.2. Installation

Follow the standard installation of python interface in DeePMD-kit. After that, you can smoothly use the DPA-1 model with following instructions.

3.5.3. Introduction to new features of DPA-1

Next we will list the detail settings in input.json and the data format, especially for large systems with dozens of elements. An example of DPA-1 input can be found in here. Descriptor "se_atten"

The notation of se_atten is short for the smooth edition of Deep Potential with an attention mechanism. This descriptor was described in detail in the DPA-1 paper and the images above.

In this example we will train a DPA-1 model for a water system. A complete training input script of this example can be find in the directory:


With the training input script, data are also provided in the example directory. One may train the model with the DeePMD-kit from the directory.

An example of the DPA-1 descriptor is provided as follows

	"descriptor" :{
          "type":		"se_atten",
          "rcut_smth":	0.50,
          "rcut":		6.00,
          "sel":		120,
          "neuron":		[25, 50, 100],
          "axis_neuron":	16,
          "resnet_dt":	false,
          "attn":	128,
          "attn_layer":	2,
          "attn_mask":	false,
          "attn_dotr":	true,
          "seed":	1
  • The type of the descriptor is set to "se_atten", which will use DPA-1 structures.

  • rcut is the cut-off radius for neighbor searching, and the rcut_smth gives where the smoothing starts.

  • sel gives the maximum possible number of neighbors in the cut-off radius. It is an int. Note that this number highly effects the efficiency of training, which we usually use less than 200. (We use 120 for training 56 elements in OC2M dataset)

  • The neuron specifies the size of the embedding net. From left to right the members denote the sizes of each hidden layer from input end to the output end, respectively. If the outer layer is of twice size as the inner layer, then the inner layer is copied and concatenated, then a ResNet architecture is built between them.

  • The axis_neuron specifies the size of submatrix of the embedding matrix, the axis matrix as explained in the DeepPot-SE paper

  • If the option resnet_dt is set to true, then a timestep is used in the ResNet.

  • seed gives the random seed that is used to generate random numbers when initializing the model parameters.

  • attn sets the length of hidden vector during scale-dot attention computation.

  • attn_layer sets the number of layers in attention mechanism.

  • attn_mask determines whether to mask the diagonal in the attention weights and False is recommended.

  • attn_dotr determines whether to dot the relative coordinates on the attention weights as a gated scheme, True is recommended. Fitting "ener"

DPA-1 only support "ener" fitting type, and you can refer here for detail information. Type embedding

DPA-1 only support models with type embeddings on. And the default setting is as follows:

            "neuron":           [8],
            "resnet_dt":        false,
            "seed":             1

You can add these settings in input.json if you want to change the default ones, see here for detail information. Type map

For training a large systems, especially those with dozens of elements, the type determines the element index of training data:

"type_map": [

which should include all the elements in the dataset you want to train on.

3.5.4. Data format

DPA-1 supports the standard data format, which is detailed in data-conv.md and system.md. Note that in this format, only those frames with the same fingerprint(i.e. the number of atoms of different element) can be put together as a unit system. This may lead to sparse frame number in those rare systems.

An ideal way is to put systems with same total number of atoms together, which is the way we trained DPA-1 on OC2M. This system format, which is called mixed_type, is proper to put frame-sparse systems together and is slightly different from the standard one. Take an example, a mixed_type may contain the following files:


This system contains Nframes frames with the same atom number Natoms, the total number of element types contained in all frames is Ntypes. Note that we put all the frames in one set set.000. Most files are the same as those in standard format, here we only list the distinct ones:








Atom type indexes (place holder)




All zeros to fake the type input


Atom type names




Atom names that map to atom type contained in all the frames, which is unnecessart to be contained in the periodic table


Atom type indexes of each frame



Nframes * Natoms

Integers that describe atom types in each frame, corresponding to indexes in type_map

With these edited files, one can put together frames with the same Natoms, instead of the same formula (like H2O). Note that this mixed_type format only supports se_atten descriptor.

The API to generate or transfer to mixed_type format will be uploaded on dpdata soon for a more convenient experience.

3.5.5. Training example

Here we upload the AlMgCu example showed in the paper, you can download here: Baidu disk; Google disk.