Machine-learned potentials in molecular simulation: best practices and tutorials
Since the seminal work of Behler and Parrinello in 2007,1 machine-learned potentials in molecular simulations have developed into a vibrant research field with new developments and different applications coming at an astonishing rate. Progress has been made in the design of new molecules and materials,2,3 the simulation of the movement of molecules and materials4-8 or in finding the solution to the Schrödinger equation. Much progress has been driven by the development of new descriptors to represent molecules and materials in 1D, 2D, and 3D.12-15 A multitude of programs, tools or interfaces between machine learning and traditional theoretical programs like classical force fields, molecular dynamics programs7,16,17 or quantum chemical programs to, e.g., accelerate the self-consistent field convergence,18 have been described and are still in active development19,20 to facilitate research in this field.21
The goal of this workshop is to bring together a small group of people, who have on one side contributed significantly to the advancement of machine learning in the aforementioned areas and on the other side work in research areas that have not yet or have hardly been touched by machine learning. This teaming up of researchers should lead to lively discussions and the identification of best-practices and tutorials to make machine learning more accessible to other chemistry research fields and researchers who enter this field. We suggest to use the concepts developed by the Living Journal of Computational Molecular Science (LiveCoMS; https://livecomsjournal.org/index.php/livecoms/), which provides a peer-reviewed home for manuscripts that share best practices in molecular modeling and simulation. LiveCoMS publishes living documents, which are regularly updated with community input, which is excellently suited for a dynamic research field as machine learning.
The aim at the workshop is to actively work together in small groups to formulate best practices and tutorials in topics like:
“ML basics”,15,22 where we aim to provide best practices and tutorials for machine learning training, validation, and also data curation – one of the most critical tasks in any machine learning study. We will bring together scientists and perspectives from the field of molecular chemistry and material science.
“ML/MM molecular dynamics”,6,7,23 “ML molecular dynamics in the ground state”,4,5 “ML photodynamics”,8,24 and “ML quantum dynamics” 27,28 which all focus on the investigation of the movement of molecules and materials. These topics will be discussed independently from each other with experts in each field initially to subsequently identify common ground and areas.
“ML for molecular materials design”,2,25,26 with a focus on unsupervised generative learning. Generative models mainly focus on molecular systems, but often face difficulties when being applied to materials,2 which will be discussed with machine learning26,29-31 and material science experts.32,33
By bringing together experts in the field and less-experienced researchers, we hope to lay the foundation for several such manuscripts and future collaborations that can be kept alive and updated in a community effort over the course of the next couple of years. This can be a general description of the best practices in ML and concrete tutorials for different kind of models (for example ANI,34 kernel methods,15,22,35 SchNet,13,16,17 equivariant representations12,14,36 – among others).
Chris Oostenbrink (University of Natural Resources and Life Sciences, Vienna (BOKU)) - Organiser
Julia Maria Westermayr (Leipzig University) - Organiser