Updates to i-PI package improve performance in widely used atomistic simulation software

Simulating complex systems from first principles has experienced a transformation in the past decade, largely driven by machine learning (ML). A new way of obtaining intermolecular energies is now possible thanks to the ability to extrapolate the accurate results obtained from reference electronic structure calculations. Such advances, however, require a low computational overhead for codes that simulate nuclear motion, such as the i-PI package.

Researchers have announced a new release of i-PI that lowers its computational overhead and adds new exciting features. With contributions from the i-PI community, Litman et al.’s improvements make i-PI up to a factor 10 faster, allowing efficiently integrated ML potentials into advanced simulation protocols.

“I-PI pioneered the idea of having a client-server model for running complicated simulations driven by electronic-structure calculations,” said author Mariana Rossi.

The use of ML potentials to sidestep the solution of the electronic Schrödinger equation enables (thermo)dynamic simulations of tens of thousands of atoms with i-PI. The benchmarks show this capability using widely adopted ML interatomic potentials such as the Behler-Parinello, DeePMD, and MACE neural networks.

“These developments make i-PI’s modular design compatible with the needs of state-of-the-art, ML-driven atomistic simulations,” said author Michele Ceriotti.

The team added several new features in this release, including an efficient algorithm to model bosonic and fermionic exchange, an infrastructure to handle error estimates from ML potentials, a closer integration with electronic-structure quantities, and simulations of photon-nuclear dynamics in optical or plasmonic cavities.

The team hopes the new release will encourage more researchers to use and contribute to i-PI. Future improvements look to better integrate the engine with different communication protocols and leverage GPU architectures to accelerate advanced simulation protocols.

Source: “i-PI 3.0: A flexible and efficient framework for advanced atomistic simulations,” by Yair Litman, Venkat Kapil, Yotam M. Y. Feldman, Davide Tisi, Tomislav Begušić, Karen Fidanyan, Guillaume Fraux, Jacob Higer, Matthias Kellner, Tao E. Li, Eszter Sarolta Pós, Elia Stocco, George Trenins, Barak Hirshberg, Mariana Rossi, and Michele Ceriotti, Journal of Chemical Physics (2024). The article can be accessed at https://doi.org/10.1063/5.0215869 .

This paper is part of the Modular and Interoperable Software for Chemical Physics Collection, learn more here .