Skip to main content
people in digital environment

Enabling AI for Science: Background and Motivation

There is increasing recognition among researchers, governments and science funders that AI offers the potential to shift the paradigm of scientific discovery, enabling researchers to analyze vast datasets and simulate complex phenomena at unprecedented speeds. But there are a number of practical challenges and barriers-to-entry that must be addressed if we are to realize this potential.

For increased uptake of AI in Science we need both:

  1. Understanding of how AI models can be usefully applied to a research question
  2. Practical knowledge about how to deploy and execute these models with one’s own data in workflows that are situated in complex IT ecosystems.

In this six month N8/Bede AI4Science project we focussed primarily on the latter, a problem which in a nutshell is about enabling AI for Science.

In practice, scientific AI workflows do not exist in isolation but rather in complex technical ecosystems comprising both local and potentially cloud-based hardware and software, and High Performance Computing (HPC) platforms such as Bede and Isambard-AI, the new UK tier 1 cluster. For reasons of reproducibility, scaling and sharing, both software and data, often with complex dependencies, need to be moved between, deployed and executed on different and continuously evolving IT platforms, with the scientific requirement that identical results be produced regardless of the platform used or at what point in time the software is executed and all the while ensuring that the data is secure.

To do this is far from trivial and typically daunting and/or beyond the capability and time-resource of the discipline-focussed, individual research scientist.

To mitigate these technical hurdles and thus to facilitate the uptake of AI within scientific research, scientists can benefit from various kinds of assistance.

One is the creation of software tools that insulate the scientist from the underlying technical details, e.g. through virtualisation and containerisation, bundling common sequences of operations, and so on.

Another is human expert assistance in the form of dedicated digital research technology professionals (dRTPs) e.g. research software engineers (RSEs) who have the training and knowledge to support the adaptation, deployment and execution of AI models on disparate platforms, either through practical implementation, close collaboration with researchers and/ or through the provision of easy to follow guidelines for researchers to follow.

In this project we explored various approaches to the problem and provided dRTPs with an opportunity to lead interventions. Key project highlights included:

  • Enhanced accessibility. dRTPs were tasked with improving accessibility and encouraging wider uptake of AI in science, with interventions looking at the migration of legacy research software and custom models onto Bede/GPUs; the design of custom interfaces and portability solutions such as containers and on-boarding guides that allow "non-HPC" users to interact with powerful AI tools without needing deep command-line knowledge.
  • Building teams and leadership. The project brought together key people with AI for science expertise across multiple sites, helping to strengthen links between dRTPs at participating universities. dRTPs led the interventions, meeting regularly with project coordinators and members of the Bede support team and providing the specialist skills and know-how needed for supporting researcher access to AI workflows.
  • Best Practice. The project adopted best practice throughout, hosting key models and datasets within a secure environment, ensuring that scientific rigor, data sensitivity and ethics were not compromised and adhering to FAIR principles where possible.
  • Sustainability. We focussed on projects that could be hosted on Bede (HPC), but with longevity built in (i.e. the results could be ported easily to other newer options eg Isambard-AI the UK’s new tier 1 compute facility).

Read the Enabling AI for Science Case Studies

Return to article index