Version: 2024.05-cpu

Parameter details

This section introduces all the parameters that can be defined by the user in the models, which can be divided into necessary parameters and advanced parameters. Necessary parameters need to be specified by the user, while advanced parameters have default values that can be manually modified in the JSON file as needed. In the parameters below, "relative path" refers to the path relative to the current working directory, while "absolute path" refers to the complete path of a file or directory starting from the root directory.

Required

For any model, the following parameters need to be provided by the user.

model_type

This parameter is used to specify which model to use for training. You can use DP for deep learning, NN for neural network, LINEAR for linear models, or the NEP for NEP model.

atom_type

This parameter is used to set the element types of the training system. The atomic number of the input elements is specified by the user in any desired order. For example, for a single-element system like copper, it would be set as [29], and for multi-element systems like CH4, it would be set as [1, 6]. You could also use the name of the atom type, such as ["Cu"] or ["H", "C"].

max_neigh_num

size of neighbor buffer, with default value 100. However, for some systems it is not enough to accommodate all the neighbors, and the feature generation fails. The following warning will pop up:

Error! maxNeighborNum too small

In this case increase the value.

raw_files

This parameter specifies the path to the molecular dynamics trajectory files for the training task. You can use either a relative or absolute path.

For DP and NEP models, supported file formats include PWmat, VASP, and CP2K (with the corresponding format parameters being pwmat/movement, vasp/outcar, cp2k/md).
For LINEAR and NN models, only the pwmat/movement format is supported.

format

This parameter works in conjunction with raw_files to specify the format of the raw trajectory files, with a default value of pwmat/movement. PWmat, VASP, and CP2K correspond to format parameters of pwmat/movement, vasp/outcar, and cp2k/md, respectively.

datasets_path

This parameter is for DP and NEP models, specifying the path to the data in pwmlff/npy format. If raw_files is specified for DP and NEP models, the trajectories from raw_files will be automatically converted to pwmlff/npy format during training or testing. The conversion tool is PWDATA.

train_movement_file

In versions prior to PWMLFF-2025.5, this parameter was used to provide the movement file path for training in LINEAR and NN models. The current version (2024.5) has modified this to raw_files, which is also compatible with this parameter, meaning you can use either train_movement_file or raw_files.

test_movement_file

In versions prior to PWMLFF-2025.5, this parameter was used to provide the movement file path for testing in LINEAR and NN models. The current version (2024.5) has modified this to raw_files, which is also compatible with this parameter, meaning you can use either test_movement_file or raw_files.

model_load_file

This parameter specifies the path to the model used for the testing task.

Advanced Parameters

Users only need to set the necessary parameters to complete the training, testing, and related molecular dynamics processes. The corresponding advanced parameters, such as model hyperparameters and optimizer hyperparameters, will be set to their default values. These advanced parameters can also be configured in the JSON file.

train_valid_ratio

This parameter specifies the ratio of the training set to the validation set. For example, 0.8 indicates that the first 80% of the images in MOVEMENT will be used as the training set, with the remaining 20% as the validation set. The default value is 0.8.

recover_train

This parameter is used to resume training from an interrupted DP or NN training task. The default value is true.

work_dir

This parameter sets the working directory for executing training, testing, and other tasks. It can be set to an absolute path or a relative path. The default value is the relative path ./work_dir.

reserve_work_dir

This parameter specifies whether to keep the working directory work_dir after the task execution is completed. The default value is False, meaning the directory will be deleted after execution.

info

Note that the parameters work_dir, reserve_work_dir, train_movement_file, and test_movement_file are only applicable to LINEAR and NN models. Also, LINEAR and NN models only support movement file formats. The train_movement_file and test_movement_file parameters have been replaced by raw_files in the PWMLFF-2024.5 version, while still being compatible with these parameters.

type_embedding

Setting "type_embedding":true will use the type embedding method to train the DP model. The default value is false. (This parameter is only used in the DP model. That is "model_type":"DP")

model 参数

The complete DP model includes three components: descriptor, fitting_net, and type_embedding. The NN model does not include type_embedding. The Linear model does not require specifying the model parameter. The parameters for the NEP model need to be set separately.

Linear model

The complete parameters for the Linear model are as follows:

    "model": {
        "descriptor": {
            "Rmax": 6.0,
            "Rmin": 0.5,
            "feature_type": [3,4]
        }
    }

Rmax

The maximum truncation radius for the features. The default value is $6.0 \text{\AA}$ .

Rmin

The minimum truncation radius for the features. The default value is $0.5 \text{\AA}$ .

feature_type

The parameter is for feature type. The supported options are [1, 2], [3, 4], [5], [6], [7], and [8]. The default value is [3, 4], 2-b and 3-b Gaussian feature. For more detailed information on different feature types, please refer to Appendix 1.

NN model

The complete set of parameters for the NN model is as follows.

    "model": {
        "descriptor": {
            "Rmax": 6.0,
            "Rmin": 0.5,
            "feature_type": [3,4]
        },
        "fitting_net": {
            "network_size": [15,15,1]
        }
    }

Rmax

The maximum truncation radius for the features. The default value is $6.0 \text{\AA}$ .

Rmin

The minimum truncation radius for the features. The default value is $0.5 \text{\AA}$ .

feature_type

network_size

The parameter is for the structure of the fitting_net. The defult value are [15, 15, 1], its structure is as: Input layer (Input data dimensions) ➡ Hidden layer 1 (15 neurons) ➡ Hidden layer 2 (15 neurons) ➡ Output layer (1 neuron)

DP model

The complete set of parameters for the DP model is as follows.

    "model": {
        "type_embedding":{
            "physical_property":["atomic_number", "atom_mass", "atom_radius", "molar_vol", "melting_point", "boiling_point", "electron_affin", "pauling"]
        },
        "descriptor": {
            "Rmax": 6.0,
            "Rmin": 0.5,
            "M2": 16,
            "network_size": [25,25,25]
        },
        "fitting_net": {
            "network_size": [50,50,50,1]
        }
    }

physical_property

This parameter is used to specify the required parameters for type embedding training in the DP model. We provide eight physical properties for users to choose.

atomic_number: Atomic number atom_mass: Atomic mass atom_radius: Atomic radius molar_vol: Molar volume melting_point: Melting point boiling_point: Boiling point electron_affin: Electron affinity pauling: Pauling electronegativity

The default value of "physical_property" is ["atomic_number", "atom_radius", "atom_mass", "electron_affin", "pauling"]

Rmax

The maximum truncation radius for the features. The default value is $6.0 \text{\AA}$ .

Rmin

The minimum truncation radius for the features. The default value is $0.5 \text{\AA}$ .

M2

The parameter is for the DP model network, determining the output size of the embedding network and the input size of the fitting network. In the example, the embedding network output size is (25 * 16), and the fitting network input size is (25 * 16 = 400). The default value is 16.

network_size

The parameter is for the structure of the embedding_net and the fitting_net. The defult value are [25, 25, 25] and [50, 50, 50, 1], respectively. The corresponding network structure is described as follows. structure of the embedding_net is: Input layer (Input data dimensions) ➡ Hidden layer 1 (25 neurons) ➡ Hidden layer 2 (25 neurons) ➡ output layer (25 neurons) structure of the fitting_net is: Input layer (M2 X 25) ➡ Hidden layer 1 (50 neurons) ➡ Hidden layer 2 (50 neurons) ➡ Hidden layer 3 (50 neurons) ➡ Output layer (1 neuron)

NEP model

The complete parameter settings for the NEP model are as follows.

{
    "model_type": "NEP",
    "atom_type": [8,72],
    "max_neigh_num": 100,
    "model": {
        "descriptor": {
            "cutoff": [6.0,6.0],
            "n_max": [4,4],
            "basis_size": [12,12],
            "l_max": [4,2,1],
            "zbl": 2.0
        },
        "fitting_net": {
            "network_size": [100,1]
        }
    }
}

model_type

This parameter specifies the type of NEP training.

cutoff

This parameter sets the cutoff energies for radial and angular components. In the implementation of PWMLFF, only the radial cutoff energy is used, and the angular cutoff energy is the same as the radial cutoff energy. The default value is [6.0, 6.0].

n_max

This parameter sets the number of features for the distances and angles corresponding to radial and angular components, respectively. The default value is [4, 4].

basis_size

This parameter sets the number of basis functions for the distances and angles corresponding to radial and angular components, respectively. The default value is [12, 12].

l_max

This parameter sets the expansion order for angular components and also controls whether four-body and five-body features are used. The default value is [4, 2, 1], corresponding to the orders for three-body, four-body, and five-body features, respectively. Here, 2 indicates the use of four-body features, and 1 indicates the use of five-body features. If you only use three-body features, set this to [4, 0, 0]; if you only use three-body and four-body features, set this to [4, 2, 0].

network_size

This parameter sets the number of neurons in the hidden layer of the NEP model. The NEP model has only one hidden layer by default, with the default value being [100]. Although multi-layer neural networks are supported (e.g., you can set it to [50, 50, 50, 1]), we recommend using the default value. In our tests, adding more network layers did not significantly improve model fitting accuracy and instead increased inference burden, reducing inference speed.

zbl

This parameter is used to set the Ziegler-Biersack-Littmark (ZBL) potential, which handles situations where atomic distances are extremely close. By default, it is not set. The allowed range for this value is 1.0 $\le$ zbl $\le$ 2.5.

optimizer

The optimizers available for training DP, NRP, NN models are the KF(Kalman Filter) Optimizer and ADAM optimizer.

KF optimizer

KF optimizer's complete parameter settings are as follows:

    "optimizer": {
        "optimizer": "LKF",
        "epochs": 30,
        "batch_size": 1,
        "print_freq": 10,
        "block_size": 5120,
        "p0_weight": 0.01,
        "kalman_lambda": 0.98,
        "kalman_nue": 0.9987,
        "train_energy": true,
        "train_force": true,
        "train_ei": false,
        "train_virial": false,
        "train_egroup": false,
        "pre_fac_force": 2.0,
        "pre_fac_etot": 1.0,
        "pre_fac_ei": 1.0,
        "pre_fac_virial": 1.0,
        "pre_fac_egroup": 0.1
    }

optimizer

This parameter is used to specify the optimizer name, and the available options are LKF or GKF. For more detailed information on the optimizer, please refer to the article, which provides more in-depth details about the implementation and characteristics of the optimizer.

epochs

This parameter is used to specify the number of training epochs. In machine learning, one epoch refers to a complete pass of the entire training dataset through the neural network, including both forward propagation and backward propagation. During each epoch, the training dataset is divided into mini-batches of samples, and each batch is input to the neural network for forward propagation, loss calculation, and parameter updates through backward propagation. The number of epochs determines how many times the entire training dataset will be processed during the training process. The dedault value is 30.

Choosing the appropriate number of epochs is typically done through debugging and evaluation during the training process. If the number of epochs is too small, the model may not learn the patterns and features of the dataset sufficiently, resulting in underfitting. On the other hand, if the number of epochs is too large, the model may overfit the training data, leading to a decrease in generalization performance on new data. Therefore, selecting an appropriate number of epochs is crucial for training an effective neural network model.

batch_size

The parameter of batch size determines the number of training samples included in each mini-batch during an epoch. The dedault value is 1.

print_freq

The parameter for specifying how often to print the training error after a certain number of mini-batch iterations. The default value is 10.

block_size

The parameter is a hyperparameter for the LKF optimizer, which specifies the block size of the covariance matrix P. A larger block size increases memory and GPU memory consumption, leading to slower training, while a smaller block size affects convergence speed and accuracy. The default value is 5120. If using high-end graphics cards such as A100 or H100, it is recommended to set it to 10240.

p0_weight

This parameter is a hyperparameter for the LKF and GKF optimizers, used for regularization. By default, it is 0.01, meaning regularization is applied. Setting a regularization term helps reduce model overfitting. The value for this parameter must be less than 1.0, with 0.01 being found as a suitable value through testing.If set to 1, it means regularization is not applicable.

kalman_lambda

The parameter is a hyperparameter for the LKF and GKF optimizer. it is called memory factor. The greater it is, the more weight, or say attention, is paid to previous data. The default value is 0.98.

kalman_nue

The parameter is a hyperparameter for the LKF and GKF optimizer. kalman_nue is forgetting rate, a hyperparameter describing the varying rate of kalman_lambda. The default value is 0.9987.

train_energy

The parameter is used to specify whether to train the total energy, The default value is true.

train_force

The parameter is used to specify whether to train the total energy, The default value is true.

train_ei

The parameter is used to specify whether to train the atomic energy, The default value is false.

train_virial

The parameter is used to specify whether to train the virial, The default value is false.

train_egroup

The parameter is used to specify whether to train the total energy, The default value is false.

pre_fac_etot

This parameter is used to specify the weight or contribution of the total energy to the loss. The default value is 1.0.

pre_fac_force

This parameter is used to specify the weight or contribution of the force to the loss. The default value is 2.0.

pre_fac_ei

This parameter is used to specify the weight or contribution of the atomic energy to the loss. The default value is 1.0.

pre_fac_virial

This parameter is used to specify the weight or contribution of the virial to the loss. The default value is 1.0.

pre_fac_egroup

This parameter is used to specify the weight or contribution of the egroup to the loss. The default value is 0.1.

ADAM optimizer

The complete set of parameters for the ADAM optimizer is as follows.

    "optimizer": {
        "optimizer": "ADAM",
        "epochs": 30,
        "batch_size": 1,
        "print_freq": 10,
        "lambda_2": 0.1,
        "learning_rate": 0.001,
        "stop_lr": 3.51e-08,
        "stop_step": 1000000,
        "decay_step": 5000,
        "train_energy": true,
        "train_force": true,
        "train_ei": false,
        "train_virial": false,
        "train_egroup": false,
        "start_pre_fac_force": 1000,
        "start_pre_fac_etot": 0.02,
        "start_pre_fac_ei": 0.1,
        "start_pre_fac_virial": 50.0,
        "start_pre_fac_egroup": 0.02,
        "end_pre_fac_force": 1.0,
        "end_pre_fac_etot": 1.0,
        "end_pre_fac_ei": 2.0,
        "end_pre_fac_virial": 1.0,
        "end_pre_fac_egroup": 1.0
    }

optimizer, epochs, batch_size, print_freq, train_energy, train_force, train_ei, train_virial, train_egroup. These parameters have the same functionality in the KF optimizer.

lambda_2

This parameter is used to set the L2 regularization term for the Adam optimizer. By default, it is not set. Setting a regularization term helps reduce model overfitting.

learning_rate

The parameter is the initial learning rate for the Adam optimizer. The default value is 0.001.

stop_lr

The parameter refers to the stopping learning rate, indicating that the training process will stop when the learning rate decreases to that value. The default value is 3.51e-08.

stop_step

The parameter refers to the stopping step, indicating that the training process will stop when reaching that step. The default value is 1000000.

decay_step

This parameter represents the decay step, which indicates the interval at which the learning rate is decayed. After each decay step, the learning rate is updated according to a certain decay rate. The default value is 5000.

learning_rate, stop_lr, stop_step, decay_step These four variables are used to update the learning rate, and the computation process is shown below in the following Python code or mathematical formula:

decay_rate = np.exp(np.log(stop_lr/learning_rate) / (stop_step/decay_step))
real_lr = learning_rate * np.power(decay_rate, (iter_num//decay_step))

The calculation formula for the decay_rate.：

\text{decay\_rate} = \exp\left(\frac{\log(\text{stop\_lr}/\text{start\_lr})}{\text{stop\_step}/\text{decay\_step}}\right)

The calculation formula for the real learning rate：

\text{real\_lr} = \text{start\_lr} \cdot \left(\mathrm{decay\_rate}\right)^{\left(\left\lfloor\frac{\text{iter\_num}}{\text{decay\_step}}\right\rfloor\right)}

Where iter_num represents the number of iterations in the training process.

start_pre_fac_force

The prefactor of force loss at the start of the training, should be larger than or equal to 0. The default value is 1000.

start_pre_fac_etot

The prefactor of total energy loss at the start of the training, should be larger than or equal to 0. The default value is 0.02.

start_pre_fac_ei

The prefactor of atomic energy loss at the start of the training. Should be larger than or equal to 0. The default value is 0.1.

start_pre_fac_virial

The prefactor of virial loss at the start of the training. Should be larger than or equal to 0. The default value is 50.0.

start_pre_fac_egroup

The prefactor of egroup loss at the start of the training, Should be larger than or equal to 0. The default value is 0.02.

end_pre_fac_force

The prefactor of force loss at the end of the training, Should be larger than or equal to 0. The default value is 1.0.

end_pre_fac_etot

The prefactor of total energy loss at the end of the training, should be larger than or equal to 0. The default value is 1.0.

end_pre_fac_ei

The prefactor of atomic energy loss at the end of the training, Should be larger than or equal to 0. The default value is 2.0.

end_pre_fac_virial

The prefactor of virial loss at the end of the training, Should be larger than or equal to 0. The default value is 1.0.

end_pre_fac_egroup

The prefactor of egroup loss at the end of the training, Should be larger than or equal to 0. The default value is 1.0.

Parameter details

Required​

model_type​

atom_type​

max_neigh_num​

raw_files​

format​

datasets_path​

train_movement_file​

test_movement_file​

model_load_file​

Advanced Parameters​

train_valid_ratio​

recover_train​

work_dir​

reserve_work_dir​

type_embedding​

model 参数​

Linear model​

Rmax​

Rmin​

feature_type​

NN model​

Rmax​

Rmin​

feature_type​

network_size​

DP model​

physical_property​

Rmax​

Rmin​

M2​

network_size​

NEP model​

model_type​

cutoff​

n_max​

basis_size​

l_max​

network_size​

zbl​

optimizer​

KF optimizer​

optimizer​

epochs​

batch_size​

print_freq​

block_size​

p0_weight​

kalman_lambda​

kalman_nue​

train_energy​

train_force​

train_ei​

train_virial​

train_egroup​

pre_fac_etot​

pre_fac_force​

pre_fac_ei​

pre_fac_virial​

pre_fac_egroup​

ADAM optimizer​

lambda_2​

learning_rate​

stop_lr​

stop_step​

decay_step​

start_pre_fac_force​

start_pre_fac_etot​

start_pre_fac_ei​

start_pre_fac_virial​

start_pre_fac_egroup​

end_pre_fac_force​

end_pre_fac_etot​

end_pre_fac_ei​

end_pre_fac_virial​

end_pre_fac_egroup​

Required

model_type

atom_type

max_neigh_num

raw_files

format

datasets_path

train_movement_file

test_movement_file

model_load_file

Advanced Parameters

train_valid_ratio

recover_train

work_dir

reserve_work_dir

type_embedding

model 参数

Linear model

Rmax

Rmin

feature_type

NN model

Rmax

Rmin

feature_type

network_size

DP model

physical_property

Rmax

Rmin

M2

network_size

NEP model

model_type

cutoff

n_max

basis_size

l_max

network_size

zbl

optimizer

KF optimizer

optimizer

epochs

batch_size

print_freq

block_size

p0_weight

kalman_lambda

kalman_nue

train_energy

train_force

train_ei

train_virial

train_egroup

pre_fac_etot

pre_fac_force

pre_fac_ei

pre_fac_virial

pre_fac_egroup

ADAM optimizer

lambda_2

learning_rate

stop_lr

stop_step

decay_step

start_pre_fac_force

start_pre_fac_etot

start_pre_fac_ei

start_pre_fac_virial

start_pre_fac_egroup

end_pre_fac_force

end_pre_fac_etot

end_pre_fac_ei

end_pre_fac_virial

end_pre_fac_egroup