PhysNet

Overview

PhysNet provides a TensorFlow implementation of the neural network described in the linked PhysNet paper. The README identifies energies, forces, dipole moments and partial charges as the model’s prediction targets through the cited paper title. Its documented workflow centers on training a model using a prepared dataset and user-selected settings, rather than accessing a hosted prediction service.

Training is configured through config.txt, where users specify hyperparameters, the dataset location and the sizes of the training and validation sets. The README directs users to train.py for the full list of available options and identifies that file as the training entry point. The supplied configuration expects sn2_reactions.npz, an example dataset available through the linked Zenodo record. Alternative datasets must follow the same format; the dataset’s own README is the designated reference for those formatting details.

The stated runtime requirements are Python 3 and TensorFlow. The README reports testing with Python 3.6.3 and TensorFlow 1.10.1, but the supplied evidence does not establish compatibility with other versions. It also does not describe the exact input fields, prediction-file formats, pretrained weights or an inference procedure. These gaps matter when planning integration into a molecular-property workflow: the excerpts support a configurable research training implementation, but do not establish a ready-to-use prediction interface or measured accuracy for a particular application.

Key Features

• TensorFlow implementation of the PhysNet neural network described in the linked research paper.
• Model prediction targets identified as energies, forces, dipole moments and partial charges.
• Training configuration through config.txt, including hyperparameters, dataset location and training/validation set sizes.
• Training entry point in train.py, which also documents the available configuration options.
• Default configuration for sn2_reactions.npz, with support for alternative datasets that follow the example’s format.

Use Cases

• Intended evaluation: assess the documented training workflow using the linked sn2_reactions.npz dataset.
• Intended evaluation: prepare a molecular dataset in the example format and investigate PhysNet for the property targets identified in the cited paper.
• Intended evaluation: explore how documented hyperparameter and training/validation split settings affect a research training workflow.

How to Use

1. Read the official README to understand the training workflow and stated requirements. It reports testing with Python 3.6.3 and TensorFlow 1.10.1; other environments need separate compatibility assessment.
2. Consult the PhysNet paper for model details before deciding whether its prediction targets match your research task.
3. Obtain the example dataset from the Zenodo record. Read that dataset’s README for formatting requirements, especially if you intend to substitute your own data.
4. In the repository, inspect train.py for available options and edit config.txt to set the dataset location, hyperparameters and training/validation sizes.
5. Start training through train.py as directed by the README. Treat the resulting run as an evaluation: inspect its outputs and assess suitability for your intended task rather than assuming accuracy or transferability.