NNsForMD

Neural network class for molecular dynamics to predict potential energy, gradients and non-adiabatic couplings (NACs).

Table of Contents

• General
• Installation
• Documentation
• Usage
• Examples
• Citing
• References

General

This repo is currently under construction. The original version used as the PyRAI2MD interface is v1.0.0.

Installation

Clone repository https://github.com/aimat-lab/NNsForMD and install for example with editable mode:

pip install -e ./pyNNsMD

or latest release via Python Package Index.

pip install pyNNsMD

Documentation

Auto-documentation generated at https://pynnsmd.readthedocs.io/en/latest/index.html

Usage

Ensemble

The main class pyNNsMD.NNsMD.NeuralNetEnsemble holds a list of keras models and custom scaler classes to transform or standardize input/output. Construction of NeuralNetEnsemble requires a filepath and the number of model instances to keep.

from pyNNsMD.NNsMD import NeuralNetEnsemble
nn = NeuralNetEnsemble("TestEnergy/", 2)

Adding the models and scaler classes to NeuralNetEnsemble via create. Custom classes can be added to the modules in pyNNsMD.models and pyNNsMD.scalers, but which must implement proper config and weight handling. Note that data format between model and scaler must be compatible. Instead of class instances a deserialization via keras config-dictionaries is supported for create().

from pyNNsMD.models.mlp_e import EnergyModel
from pyNNsMD.scaler.energy import EnergyStandardScaler

nn.create(models=[EnergyModel(atoms=12, states=2), EnergyModel(atoms=12, states=2)],
          scalers=[EnergyStandardScaler(), EnergyStandardScaler()])

The models and scaler must be saved to disk to prepare for training, which includes config and weights.

nn.save()

Data

The data is stored to the directory specified in NeuralNetEnsemble. Data format passed to NeuralNetEnsemble.data() must be nested python-only lists. The geometries are stored as .xyz and everything else as .json. Note that the training scripts must be compatible with the data format.

atoms = [["C", "C"]]
geos = [[[0.147, 0.024, -0.680], [-0.165, -0.037, 0.652]]]
energy = [[-20386.37, -20383.93]]

nn.data(atoms=atoms, geometries=geos, energies=energy)
# nn.data_path("data_dir/") if data can't be saved in working directory.

Training

For training the train and test indices must also be saved to file for each model directory. This can be achieved via train_test_split(), or by directly passing an index-list for each model with train_test_indices(). Note that the different models are sought to be trained on different splits.

nn.train_test_split(dataset_size=1, n_splits=1, shuffle=True) # Usually n_splits=5 or 10
# nn.train_test_indices(train=[np.array([0]), np.array([0])], test=[np.array([0]), np.array([0])])

The hyperparameter for training are passed as .json to each model folder. See pyNNsMD.hypers modules for example hyperparameter.

nn.training([{
    'initialize_weights': True, 'epo': 1000, 'batch_size': 64, 'epostep': 10, 
    'learning_rate': 1e-3, "callbacks": [], 'unit_energy': "eV", 'unit_gradient': "eV/A"
}]*2, fit_mode="training")

Fitting

With fit() a training script is run for each model from the model's directory. The training script should be stored in pyNNsMD.training. Note that the training script must be compatible with model and data. The training script must provide command line arguments 'index', 'filepath', 'gpus' and 'mode'. The training can be distributed on multiple or a single gpu (for small networks).

fit_error = nn.fit(["training_mlp_e"]*2, fit_mode="training", 
                   gpu_dist=[0, 0], proc_async=True)
print(fit_error)

Loading

After fitting the model can be recreated from config and the weights loaded from file with load().

nn.load()

Prediction

The model's prediction can be obtained from the corresponding input data via predict() and call(). The both input and output is rescaled by the scaler to match the model standardized input and output. Furthermore, the subclassed model should implement call_to_tensor_input and call_to_numpy_output or optionally predict_to_tensor_input and predict_to_numpy_output, if the model requires a specific tensor input as in call() and the scaler class usually works on numpy data.

test = nn.predict(geos)
# test_batch = nn.call(geos[:32])  # Faster than predict.

Examples

A set of examples can be found in examples, that demonstrate usage and typical tasks for projects.

Citing

If you want to cite this repository or the PyRAI2MD code, please refer to our publication at:

@Article{JingbaiLi2021,
    author ="Li, Jingbai and Reiser, Patrick and Boswell, Benjamin R. and Eberhard, André and Burns, Noah Z. and Friederich, Pascal and Lopez, Steven A.",
    title  ="Automatic discovery of photoisomerization mechanisms with nanosecond machine learning photodynamics simulations",
    journal  ="Chem. Sci.",
    year  ="2021",
    pages  ="-",
    publisher  ="The Royal Society of Chemistry",
    doi  ="10.1039/D0SC05610C",
    url  ="http://dx.doi.org/10.1039/D0SC05610C"
}

References

• https://onlinelibrary.wiley.com/doi/full/10.1002/qua.24890
• https://pubs.acs.org/doi/abs/10.1021/acs.chemrev.0c00749