Choose timezone
Your profile timezone:
ML4Accelerators @Cockcroft Working Group
→
Europe/London
(next to CLARA control room) (CR1 DL)
(next to CLARA control room)
CR1 DL
,
Description
The aim of the ML4Accelerators WG is to share progress on projects, troubleshoot common issues (coding errors, file organization, data standards, repositories), and exchange practical know-how. We’ll also use the space to keep each other updated on conferences, abstract submissions and deadlines, new papers, and relevant news.
This is an informal but structured gathering: a place to discuss what you worked on last month, learn from each other’s experiences, and make sure nobody gets stuck in isolation. Whether you’re working on experiments, simulations, or coding pipelines, this is a chance to get feedback, stay on top of developments, and build good habits as a community.
| When | the first Thursday of every month | |
| Where |
In person: meeting room next to CLARA's control room (Daresbury Laboratory) |
Remote: Teams link |
| Who |
the meeting is open to students, postdocs, staff, academics from the Cockcroft Institute and anyone interested in discussing AI/ML |
|
| What |
Rountable of attendants: share your latest news and work on AI/ML (max 5 min/person). Latest news: chairperson will go through latest news, upcoming events, and interesting papers (max 10 min). |
Rest of the meeting will have different options depending on the month:
|
- Join the ML4Accelerators Microsoft group to join the mailing list, get the calendar events, and access the documents.
- Link to add topics of discussion (you need to be part of the group mlatcockroft@stfc365.onmicrosoft.com):
ML4AcceleratorsWG-meeting-schedule.xlsx - Link for guidelines on technical discussions: technical-discussion-guidelines
Mentioned today in the technical discussion
- VAE: https://arxiv.org/pdf/1312.6114
- Conditional VAE: https://arxiv.org/pdf/2102.12037
- Please also see this paper on frequency-domain loss, which can improve image reconstruction quality: https://link.springer.com/article/10.1007/s11760-021-01915-4
- https://towardsdatascience.com/conditional-variational-autoencoders-with-learnable-conditional-embeddings-e22ee5359a2a/
- https://ijdykeman.github.io/ml/2016/12/21/cvae.html
Notes from Amy:
Instead of training an autoencoder to learn a latent space, and then a separate prediction network that maps from the parameters to that latent space, you can train the whole thing as a Conditional Variational Autoencoder. Essentially this works like training a standard autoencoder, but you feed the machine parameters into both the encoder and decoder sections, and add a constraint to your latent space so that it comes to represent a probability distribution conditioned on those parameters. By using KL divergence as a regulariser on your latent space, you force the elements of the latent space to represent a smooth and continuous probability distribution. You then concatenate the parameters to that latent space and feed it into your decoder, such that it learns to reconstruct samples from that conditioned probability distribution. This lets you train end-to-end a single network with all the properties of your two networks, as well as a latent space that you can do interesting analysis on!
Events
| Acronym and full name with link | Date | Location | Comment/Info |
| 1-7 Sept 2026 |
Crete (Greece) |
Large physics community (particle, astro) talking about differentiable simulations | |
| 3rd International Reinforcement Learning Bootcamp | 16-18 Sept 2026 |
Salzburg (Austria) |
Good opportunity for students to learn RL |
| AISSAI Hackathon on AI for Particle Accelerators | 11-16 Oct 2026 |
Les domaines de Gaston (France) |
Amy will be a team leader, good opportunity for students |
|
UK Accelerator Institute Seminar "AI for particle accelerators: towards autonomous operation" |
29 Oct 2026 |
Merrison Lecture Theatre (Daresbury Lab) |
Talk by Andrea |
| Reinforcement Learning for Experimental Sciences Workshop | 12 December | Paris (France) | Simon Hirlaender will be presenting new ideas for RL in particle accelerators. Event part of NeurIPS2026. |
| Artificial Intelligence for the Electron Ion Collider (AI4EIC) 2026 | 15-18 December | Durham, North Carolina (United States) | |
|
USPAS winter school on "Optimization, Machine Learning and Artificial Intelligence for Accelerators" |
25 Jan - 5 Feb 2027 | Albuquerque (United States) | Andrea will be teaching the RL part. Great opportunity for an in-depth ML course!! |
| RL4AA'27 (5th collaboration workshop on Reinforcement Learning for Autonomous Accelerators) | 21-23 April 2027 | SLAC (United States) | |
| MaLAPA 2027 (7th Beam Dynamics Mini-Workshop on Machine Learning for Particle Accelerators) | 26-29 April 2027 | Berkley (United States) |
Interesting/recent papers
- Reinforcement Learning applied to Optimization of LHC beams in the CERN Proton Synchrotron (Joel Wulff et al)
- Autonomous discovery of accelerator commissioning algorithms (Thorsten Hellert)
- Flow-based surrogate models for particle tracking (Matthias Remta et al)
-
Recurrent neural network encoder–decoder surrogate models for replacing computational beam simulations in beamline optimization (Xi Cheng et al)
Interesting news
There are minutes attached to this event.
Show them.