hide
Free keywords:
Physics, Chemical Physics, physics.chem-ph, Condensed Matter, Materials Science, cond-mat.mtrl-sci
Abstract:
Atomic-scale simulations have progressed tremendously over the past decade, largely due to the availability of interatomic potentials based on machine-learning algorithms. These potentials enable the combination of the accuracy of electronic structure calculations with extensive length and time scales. In this paper, we present a new release of the i-PI code that allows the community to fully benefit from the rapid developments in the field. i-PI is a Python software that facilitates the integration of different methods and different software tools by using a socket interface for inter-process communication. The current framework is flexible enough to support rapid prototyping and the combination of various simulation techniques, while maintaining a speed that prevents it from becoming the bottleneck in most workflows. We discuss the implementation of several new features, including an efficient algorithm to model bosonic and fermionic exchange, a framework for uncertainty quantification to be used in conjunction with machine-learning potentials, a communication infrastructure that allows deeper integration with electronic-driven simulations, and an approach to simulate coupled photon-nuclear dynamics in optical or plasmonic cavities. For this release, we also improved some computational bottlenecks of the implementation, reducing the overhead associated with using i-PI over a native implementation of molecular dynamics techniques. We show numerical benchmarks using widely adopted machine learning potentials, such as Behler-Parinello, DeepMD and MACE neural networks, and demonstrate that such overhead is negligible for systems containing between 100 and 12000 atoms.