PANNA 2.0: Efficient neural network interatomic potentials and new architectures.

Area of Science:

• Computational materials science
• Artificial intelligence in chemistry
• Materials informatics

Background:

• Developing accurate interatomic potentials is crucial for molecular simulations.
• Neural network potentials offer a data-driven approach to modeling atomic interactions.
• Existing methods may lack efficiency or comprehensive features for complex systems.

Purpose of the Study:

• Introduce PANNA 2.0, an updated code for generating neural network interatomic potentials.
• Highlight new features enhancing usability, performance, and scope.
• Provide benchmarks demonstrating the accuracy and capabilities of PANNA 2.0.

Main Methods:

• Utilizes local atomic descriptors and multilayer perceptrons for potential generation.
• Features a new backend with improved network training customization and monitoring.
• Incorporates enhanced GPU support, a fast descriptor calculator, and external code plugins.
• Implements a variational charge equilibration scheme for long-range electrostatics.

Main Results:

• PANNA 2.0 offers improved tools for network training and customization.
• Enhanced GPU support and a fast descriptor calculator accelerate computations.
• The new architecture effectively models long-range electrostatic interactions.
• Benchmarks show competitive accuracy against state-of-the-art methods on diverse datasets.

Conclusions:

• PANNA 2.0 represents a significant advancement in neural network interatomic potential generation.
• The code provides a powerful and versatile tool for computational materials science.
• Its improved features and accuracy facilitate more reliable and efficient materials simulations.