Yesterday, we held the third event in the “Data Driven Materials Design” series that Keith Butler and I have co-organised with sponsorship from the Thomas Young Centre. This session focused on machine learning force fields (MLFFs).
The general concept of the series has been a themed afternoon event that combines an invited external speaker and several early career researchers. The majority of participants have been in person with plenty of time and space for extended discussions and refreshments. The informal interactions, sorely missed over the past three years, have been a highlight.
1.0
The first event (November 2022) covered materials exploration and generative models with:
- Taylor Sparks
- Zhenzhu Li
- Luis Antunes
- Alex Squires
- Stephen Bennett
2.0
The second event (January 2023) addressed materials screening and optimisation with:
- Yousung Jung
- Anthony Onwuli
- Chengcheng Xiao
- Yifan Wu
3.0
Back to the third event that happened yesterday. Xia Liang (ICL) began by setting the scene and giving an introduction to MLFFs, with a focus on how to analyze the large trajectories they enable. Philipp Schienbein (UCL) followed up with some very interesting work on oxide/water interfaces with data-driven approaches to predicting vibrational signatures.
Then, Volker Deringer (Oxford) shared his wisdom on atomic-scale machine learning. What I appreciated most was how Volker emphasised the three key components of model building: (i) reference data; (ii) representations; and (iii) regression. He covered best practices, current limitations, and on-going developments. The recent perspective on “How to validate machine-learned interatomic potentials” gives a flavour of what he presented including the carefully crafted illustrations.
The rate of progress in this area of computational materials science is astounding. I once dabbled in phase change materials such as Ge-Sb-Te (GST), but never imagined >500,000 atom (density functional theory quality) dynamic simulations. Volker’s latest work suceeds in scaling the simulations up to directly emulate the dimensions of a nano-device (arXiv, 2022).
With the rapid evolution in techniques (e.g Atomic Cluster Expansion), implementations (e.g. Allegro), and computer hardware (e.g. Ada Lovelace), this feels like only the beginning…