EIDF 2026 September Drop-in: AI-Driven Climate Science, Automated Archives and Smarter GPU Jobs

Catch up on the September EIDF drop-in, where the team showcased a range of work being delivered on EIDF. Highlights include AI-driven climate simulations that slash compute costs, an automated pipeline for extracting data from historic document scans, and practical updates and debugging tips for the GPU service.

2026-September-drop-in0mage
At the September EIDF drop-in session, the team shared a range of projects and service updates, from sustainable climate modelling to large-scale document digitisation and improvements to the GPU service.
The session covered AI-driven climate simulations through the Continents project, an automated pipeline for extracting data from over a million historic aerial photographs, and recent GPU service changes with practical debugging advice.
Watch the full recording: https://edin.ac/4dBtOOX

A step change in climate science with AI-driven simulations

Kara Moraw introduced the EPSRC-funded Continents project, a four-year collaboration between EPCC, NCAS and NCAR focused on sustainable climate and weather modelling.
Kara explained how AI models such as ECMWF’s AIFS can emulate traditional simulations at a fraction of the compute cost, running inference on a single EIDF GPU. The team quantifies energy use across the full machine learning life cycle using the Code Carbon library.
Current work focuses on fine-tuning AIFS for UK extreme rain events, using Met Office weather-warning data to establish baselines and explore energy-efficiency optimisations.
Watch this segment: https://edin.ac/4je55DY

Automating data extraction from scans of historic documents

Bianca Prodan presented a three-month project with NCAP and the US National Archives (NARA) to automate metadata extraction from over one million historic aerial-photography degree-square plots.
Bianca walked through the pipeline: image normalisation, OCR for handwritten fields, template matching for corner detection, and pixel-to-geographic coordinate mapping. The pipeline is built as modular, checkpointed components that can run in parallel per area, and achieved strong success rates on live data.
The key takeaway: work that would take a 100-person team around a year to process manually can be run through the pipeline in roughly a week, with only failed cases flagged for human review.
Watch this segment: https://edin.ac/4h9FlWJ

Why your GPU jobs get stuck: updates and debugging tips

Alistair Grant outlined recent GPU service changes, including bin-packing and topology-aware scheduling to make larger multi-GPU jobs (four, six or eight GPUs) easier to schedule. Background updates to monitoring policies, the GPU operator and the SEF operator now support more stable PVCs.
Alistair also gave a practical walkthrough of debugging jobs that won’t run: checking pods, jobs and workloads with kubectl describe, interpreting common statuses, and reviewing events. The most frequent issues are missing PVCs, image pull errors, and insufficient queue quota. If in doubt, read the documentation, try things, and contact the Helpdesk.
Watch this segment: https://edin.ac/4xu7gaa

Discussion points

The Q&A covered several useful topics, including:
  • How Code Carbon estimates emissions from hardware power usage and regional grid averages, and the value of combining this with live emissions data for a more accurate picture.
  • Confirmation that Bianca’s project used a frontier model out of the box, which performed well enough for the three-month scope; fine-tuning would be a future extension.
  • A known quirk where evicted or timed-out workloads can remain in the queue, which is on the team’s fix list. Users are encouraged to check the workload status (not just the pod or job) when a job appears suspended.
  • Upcoming GPU service plans, including a further round of Kubernetes upgrades and alignment of GPU driver versions.

Next drop-in sessions

The next EIDF drop-in sessions are scheduled for:

Stay connected

This article was published on