pyxtal_ff.models.neuralnetwork

class pyxtal_ff.models.neuralnetwork.Dataset(data, softmax, device, memory, fc, sc)

Bases: torch.utils.data.dataset.Dataset

Defined a Dataset class based on PyTorch Dataset.

Tutorial: https://pytorch.org/tutorials/beginner/data_loading_tutorial.html.

close()

class pyxtal_ff.models.neuralnetwork.NeuralNetwork(elements, hiddenlayers, activation, random_seed, batch_size, epoch, device, alpha, softmax_beta, unit, force_coefficient, stress_coefficient, stress_group, restart, path, memory)

Bases: object

Atom-centered Neural Network model. The inputs are atom-centered descriptors: BehlerParrinello or Bispectrum. The forward propagation of the Neural Network predicts energy per atom, and the derivative of the forward propagation predicts force.

A machine learning interatomic potential can developed by optimizing the weights of the Neural Network for a given system.

Parameters:

• elements (list) – A list of atomic species in the crystal system.
• hiddenlayers (list or dict) – [3, 3] contains 2 layers with 3 nodes each. Each atomic species in the crystal system is assigned with its own neural network architecture.
• activation (str) – The activation function for the neural network model. Options: tanh, sigmoid, and linear.
• random_seed (int) – Random seed for generating random initial random weights.
• batch_size (int) – Determine the number of structures in a batch per optimization step.
• epoch (int) – A measure of the number of times all of the training vectors are used once to update the weights.
• device (str) – The device used to train: ‘cpu’ or ‘cuda’.
• force_coefficient (float) – This parameter is used as the penalty parameter to scale the force contribution relative to the energy.
• stress_coefficient (float) – This parameter is used as the balance parameter scaling the stress contribution relative to the energy.
• stress_group (list of strings) – Only the intended group will be considered in stress training, i.e. [‘Elastic’].
• alpha (float) – L2 penalty (regularization) parameter.
• softmax_beta (float) – The parameters used for Softmax Energy Penalty function.
• unit (str) – The unit of energy (‘eV’ or ‘Ha’).
• restart (str) – Continuing Neural Network training from where it was left off.
• path (str) – A path to the directory where everything is saved.
• memory (str) – There are two options: ‘in’ or ‘out’. ‘in’ will use load all descriptors to memory as ‘out’ will call from disk as needed.

calculate_loss(models, batch)

Calculate the total loss and MAE for energy and forces for a batch of structures per one optimization step.

calculate_properties(descriptor, bforce=True, bstress=False)

A routine to compute energy, forces, and stress.

descriptor: list

list of x, dxdr, and rdxdr.

energy, force, stress: bool

If False, excluding the property from calculation.

energy: float

The predicted energy

forces: 2D array [N_atom, 3] (if dxdr is provided)

The predicted forces

stress: 2D array [3, 3] (if rdxdr is provided)

The predicted stress

collate_fn(batch)

Return user-defined batch.

dump_evaluate(predicted, true, filename)

Dump the evaluate results to text files.

dump_formulas(formulas, filename)

Dump the evaluate results to text files.

evaluate(data, figname)

Evaluating the train or test data set based on trained Neural Network model. Evaluate will only be performed in cpu mode.

get_descriptors_range(data)

Calculate the range (min and max values) of the descriptors corresponding to all of the crystal structures.

Parameters:data (dict) – data contains atom-centered descriptors. Returns:The ranges of the descriptors for each chemical specie. Return type:dict

get_stress_group(data)

Get every kind of stress groups if None is defined by the user.

load_checkpoint(filename=None, method=None, args=None)

Load PyTorch Neural Network models at previously saved checkpoint.

mean_absolute_error(true, predicted)

Calculate mean absolute error of energy or force.

mean_squared_error(true, predicted)

Calculate mean square error of energy or force.

normalize(data, drange, unit, norm=[0.0, 1.0])

Normalizing the descriptors to the range of [0., 1.] based on the min and max value of the entire descriptors.

Example: X.shape == [60, 10]; len(self.elements) == 2 X_norm -> {‘element1’: [40, 10], ‘element2’: [20, 10]}

Parameters:

• data (str) – The directory path to the database
• drange – The range of the descriptors for each element species.
• unit (str) – The unit of energy (‘eV’ or ‘Ha’).
• norm (tuple of floats.) – The lower and upper bounds of the normalization.

Returns:

The normalized descriptors.

Return type:

dict

plot_hist(descriptors, figname=None, figsize=(12, 16))

Plot the histogram of descriptors.

preprocess(TrainData)

Preprocess TrainData to a convenient format for Neural Network training.

r2_score(true, predicted)

Calculate the r square of energy or force.

save_checkpoint(des_info, filename=None)

Save PyTorch Neural Network models at a checkpoint.

save_weights_to_txt(des_info)

Saving the model weights to txt file.

train(TrainData, optimizer)

Training of Neural Network Potential.