What is it about?

Interatomic potentials are a method used to simulate the forces acting on a set of atoms, allowing a detailed description of materials. They can be used - among many other possibilities - to predict the atomic structure of crystals, the diffusion of atoms within them, and their temperature-induced expansion. To do so successfully, however, these potentials must have high accuracy. Methods from machine learning have begun to be used to develop cutting-edge interatomic potentials. Here we present one such potential generation scheme, the Ephemeral Data-derived Potential, which is characterised by a particular simplicity and ease of use compared to other machine-learned potentials.

Featured Image

Why is it important?

Machine-learned potentials can be used to dramatically accelerate calculations of atomic interactions, reducing the power needed to model materials on the computer. In a world increasingly concerned with the amount of energy we use, this will help reduce the environmental impact of this scientific field. It also moves the timescales accessible in simulation and experiment closer to one another. By introducing a simple machine-learned potential, we hope to encourage the adoption of this methodology in wider circles.


Ephemeral data-derived potentials are one among a growing number of methodologies which apply techniques from machine learning to the generation of interatomic potentials. These techniques have different niches in terms of the trade-offs between accuracy, efficiency, and ease of use. Our technique emphasises the generation of fast, simple, easy-to-use potentials, but while preparing this publication we discovered that its accuracy is still in many cases cutting edge. For this reason I hope our method will become widely used and help the more widespread adoption of machine-learned potentials in general.

Pascal Salzbrenner
University of Cambridge

Read the Original

This page is a summary of: Developments and further applications of ephemeral data derived potentials, The Journal of Chemical Physics, October 2023, American Institute of Physics,
DOI: 10.1063/5.0158710.
You can read the full text:



The following have contributed to this page