Webinar: Trusted and reproducible workflows for machine learnt interatomic potentials

Register here: https://us06web.zoom.us/webinar/register/WN_sFVSMWwPR3K3uC1h-DGi6w

This webinar explores recent advances in machine learnt interatomic potentials (MLIPs) that revolutionize atomistic simulations with ab initio accuracy and expanded scales, while introducing software frameworks such as janus-core, aiida-mlip, and ML-PEG for data generation, benchmarking, training, and workflow integration within the PSDI ecosystem.

Abstract

Recent advances in machine learnt interatomic potentials (MLIPs) are revolutionising atomistic simulations, enabling atomistic modelling with accuracy comparable to ab initio calculations extending significantly time and length scales. However, in order for researchers to be able to take full advantage of these advances, software frameworks are needed to facilitate data generation, scientific benchmarking, training and fine-tuning of MLIPs, as well as to enable their integration into simulation workflows to study properties of interest. To address this need, we introduce (a) janus-core, (b) aiida-mlip and (c) ML-PEG.

The main focus of this highlight will be aiida-mlip and ML-PEG. aiida-mlip is an AiiDA plugin, enabling full provenance tracking and HPC integration for workflows involving MLIPs, such as high-throughput calculations and fine-tuning workflows. ML-PEG is an ML potential usability and performance guide, providing a framework to develop, run, and visualise an automated, modular, hierarchical test suite for MLIPs. ML-PEG is highly interactive, with users able to explore the results of tests at multiple levels of detail, and customise the relative importance and scaling of individual tests according to their applications and properties of interest. How these integrate in the larger ecosystem of PSDI, like data collections would also be highlighted.

Biography

Elliott Kasoar is a research software engineer in the data-driven materials and molecular science group within STFC’s scientific computing department. As part of the PSDI Data to Knowledge pathfinder, he leads the development of digital infrastructure for machine learnt interatomic potentials.

He is also pursuing a part-time PhD in Gábor Csányi’s group at the University of Cambridge, with a current focus on developing an ML potential usability and performance guide as both a deployed service and deployable software framework.

Dr Alin-Marin Elena is a computational scientist at STFC Daresbury Laboratory, specializing in computational statistical physics, molecular dynamics and Monte Carlo methods. He contributes to open-source scientific software, including CP2K, DL_POLY (5.0), ASE, and janus-core, and leads the Data-Driven Molecular and Materials Sciences group at STFC. With a keen interest in machine learnt interatomic potentials—their generation, usage, and application to explain experimental results—as well as computational statistical mechanics of rare events, HPC, continuous integration and deployment, and user experience for scientific codes, he leads the Data to Knowledge pathfinder in PSDI. He earned his PhD in Physics from University College Dublin in 2013 under the supervision of Prof Giovanni Ciccotti and Dr Simone Meloni. Prior to and following his PhD, he worked as a computational scientist at the Irish Centre for High End Computing, where he coordinated the National Service and participated in the Intel Parallel Computing Centre program for code modernization on emerging architectures.

He is a member of the Computational Molecular and Materials Science Theme at STFC and is involved in EPSRC/MRC/BBSRC-funded CoSeC initiatives for exchanging computational knowledge and expertise through training and outreach for CoSeC, PSDI, and the Ada Lovelace Center. Notably, he serves on the organizing committee of the CCP5 Summer School (co-sponsored by CECAM) and the CaMML school (co-sponsored by PSDI and AiHUB).

Register for this webinar

Register for this webinar directly through zoom:
https://us06web.zoom.us/webinar/register/WN_sFVSMWwPR3K3uC1h-DGi6w