June meetup 12/6/2023
First monthly meetup for SciML community at DMI
This was the first meetup of SciML @ DMI. Below are notes on what we discussed. Feel free to make a pull-request to edit any inaccuracies - Leif Denby.
What’s new in ML?
- Nvidia modulus: https://github.com/NVIDIA/modulus
- non-standard loss function?
-
sketch
package: https://github.com/approximatelabs/sketch - noteable GPT+ plugin: https://noteable.io/chatgpt-plugin-for-notebook/
What do we want SciML DMI to be?
mindmap
SciML DMI
getting practal advise on using "pascal"
would like to learn about ML, have money
brainstorm practical examples, feedback
keep up to date with recent developments
pioneer new things at DMI
have fun!
go through interesting papers, show the front page
tie in new people with ML experience, support network
learn techniques being applied in group
best practice for using ML in operations?
existing projects, what ML is being used?
DEODE
DestineE
emulation PhD student, innovation fund, Phillip Arestrup DTU
data quality analysis for Danish APA
sattellite data, sea ice mapping in radar images
flood mapping, by external partners
Anna, densification of shallow ice cores, DTU Space
physical foundation of machine learning?
use of ML at data assimilation?
GPUs?
-
platform name, hardware available, cost, how to get access
- pascal.dmi.dk, 1x A100 80GB (paid for by IT during ASIP project, Tore Wulf) only available to ASIP group (Ask Tore Wulf twu@dmi.dk, Martin Koster mak@dmi.dk), cost ~ 100,000DKK
- Tore Wulf: research group is pushing for more copies of pascal.dmi.dk
- ATOS @ ECMWF: should be easy to get access (KAH)
- European Weather Cloud
- LUMI
- EuroHPC: nvidia
- upcoming project needs:
- Ruth: emulate regional climate model over two ice sheets - do we need more GPUs?
- Fabrizio (EUMETNET project)
- no GPU UWC-W
Research projects
- ASIP (Copernicus): sea ice mapping in radar images, convolutional neural networks,
- ML (EumetSat fellowship): extending assimilation of sat microwave obs over seaice, fab@dmi.dk
- PRECISE: emulate regional climate (surface mass balance model) over ice-sheets, collab with Northumberland (sea-ice caverns) - rum@dmi.dk
- RopeWalk: digitising hand-written records of observations in ship-data tables - mas@dmi.dk
- ACCORD: use of machine learning in LAM NWP (data assimulation, ensemble forecasting and everything you can think of) - Henrik
- PhD:
- TODO
Background
- SciML Leeds community: https://sciml-leeds.github.io