PiNN

PiNN implements atomic neural networks for prediction of potential energy surfaces and physicochemical properties of molecules and materials using TensorFlow.

Key Features:

• PiNet: A graph convolutional neural network variant tailored to predict potential energy surfaces and physicochemical properties of molecules and materials.
• Behler-Parrinello neural network: Implementation of the Behler-Parrinello atomic neural network for modeling atomic interactions.
• TensorFlow backend: Uses TensorFlow as the foundational framework for implementing atomic neural networks.
• Dataset testing: Evaluated on isolated small molecules, crystalline materials, liquid water, and aqueous alkaline electrolytes.
• PiNNBoard visualizer: Extraction and analysis of chemical insights derived from atomic neural networks.
• Analytical stress tensor calculations: Tools for computing analytical stress tensors for structural and mechanical analyses.
• Integration with simulation environments: Interfaces to the atomic simulation environment and a development version of the Amsterdam Modeling Suite.
• Modular architecture: Highly modularized design enabling composition and extension for development of novel atomic neural networks.

Scientific Applications:

• Molecular dynamics simulations: Providing accurate potential energy surfaces to improve force evaluations in MD simulations.
• Materials science: Modeling and prediction of properties of crystalline materials for materials discovery and characterization.
• Chemical informatics: Prediction of physicochemical properties to support compound design and property screening.

Methodology:

Implemented in TensorFlow and providing implementations of the PiNet graph convolutional neural network variant and the Behler-Parrinello neural network, with analytical stress tensor computations, PiNNBoard-based extraction/analysis, and interfaces to the atomic simulation environment and a development version of the Amsterdam Modeling Suite.