6. Useful links ?Apptainer and bede_ml-toolkit documentation
1. The Problem
The emergence of the Matbench Discovery leaderboard has provided a vital interactive resource for ranking ML models of chemical potential based on their ability to discover stable inorganic crystals. However, a significant barrier to entry exists: every model typically requires its own isolated software stack, specific library versions, and complex configuration.
For researchers, setting up these environments is at best tedious and at worst a source of non-reproducible errors. This problem is magnified when validating research, as comparing multiple models on a single dataset requires managing several incompatible software environments simultaneously. These technical hurdles often divert specialist time away from scientific inquiry toward troubleshooting infrastructure.
2. The Solution
This case study uses software containers to streamline the deployment and accessibility of these models on the Bede HPC via Apptainer. Software containers serve as isolated layers that encapsulate applications along with their specific dependencies, ensuring they remain decoupled from the underlying system and other software environments. Instructions on how to do this can be found here.
However, a model on its own is not that useful. Instead, these models need to be integrated into other software that use these models to calculate properties of materials. Therefore, this solution provides a single entry point into the well established CASTEP code base for the models from the Matbench Discovery. This allows a user to choose a model from the Matbench leaderboard, port the model into CASTEP without having to worry about the models dependencies or having a deep understanding of CASTEP’s inner workings.
Figure 1: A model is chosen from the leaderboard. This is then wrapped in a container and ported directly into CASTEP. From there, a variety of material property calculations can be made using the model.
3. Why it helps AI4Science
Model leaderboards like Matbench Discovery are really important for benchmarking model performance; but what is the point if these models aren’t used? The work in this intervention simplifies the process of model integration in workflows and reduces setup time and therefore should free up researcher time and encourage more models to be used. Crucially, the software described in this case study not only works on the task of installing a model, but also its integration into associated software packages in the molecular dynamics field.
This kind of work supports the use of AI models for progressing science, rather than simply work to promote the position of an AI model higher up in a leaderboard or to improve the state-of-the art by an incremental amount.
4. What value did the dRTP bring?
Digital Research Technical Professionals (dRTPs) act as the vital bridge between complex infrastructure and scientific application. For this case study, the dRTP Ben Thorpe provided the specialist knowledge required to:
Implement Apptainer workflows specifically tuned for the Bede HPC environment.
In this project Ben implemented and tested two families of models, from Meta and NequIP - at least 22 different models.
Leverage deep domain expertise in DFT, specifically as a former maintainer of CASTEP, to ensure the interface is scientifically rigorous and immediately useful to the research community.
Both of these tasks require advanced software development skills that may be beyond the average researcher. Specifically, containers are a difficult, abstract concept that have a steep learning curve but are the only practical way of deploying multiple models with different dependencies.
This is a really useful intervention. There is a lot of interest in the materials modelling community in using foundation models like these to accelerate the time to science and enable quantitatively new types of investigation including high throughput materials discovery. The barrier to doing this has been the complexity of integrating these models into existing tools and workflows, and managing the many dependencies that come with each model. This intervention reduces this barrier dramatically, and enables CASTEP users to take a big step forward. Thank you Ben!
Matt Probert, Professor of Computational Physics at the University of York and AI4Science PI
