2025-03-06 Meeting notes

Date:06 Mar 2025

Attendees	Regrets	Notified
Mark Lacy

COSMIC AI

NRAO PI Eric Murphy
Brian Mason working on automating calibration
Ryan Loomis working on image analysis/image segmentation of emission in molecular line cubes.
Seminar series - next week (12th) ML will present on what we have been talking about in this NRAO meeting series. https://www.cosmicai.org/about
Can sign up for email list, Slack etc by following "Get Involved".
Call for seed funding (~$30k) proposals now open (abstracts due March 14th).

Uses of AI/ML in radio interferometric imaging

Clean works pretty well and is fast and becoming more sophisticated (e.g. ASP clean), most of the time is spent gridding the major cycles, which AI/ML probably can't help much with.
Some possible uses for AI maybe?
- "Learning" deconvolution (e.g. POLISH)
  - Preshanth and Brian Kirk have got it running, results not too impressive so far (may need more time to train)
- Preview imaging (e.g. teach AI how to fill in gaps in the UV plane, Schmidt et al. 2022) - maybe useful if processing resources limited for full processing?
  - Dillon: previews would be good, especially e.g. for time domain, want to know in ~30min if object is detected or not, not ~1 week. Again, this method quite specific (to VLBI jets), unclear how well it would do generally.
- Setting a prior in a regularized ML imaging analysis
  - Also looked at R2D2 (e.g.https://arxiv.org/abs/2503.02554) also limited (fixed grid size etc)
- Dillon and Hendrik are working on an adversarial network, train one network on calibration and imaging, other network adds artifacts (RFI etc) and then first network tries to fix the problem.

Security/privacy issues
- Some concern that using AI could be blanket banned, need to explain to management how important these new techniques are for NRAO. Agreed that some things need to be isolated (maybe on a separate network) so there is no risk of personal/private/proprietary information being used for AI. Code needs to be cleaned of information that could be used to compromise systems.
- Using web interface to help with coding OK, bit more concerning if you open up your entire repository, even if open source, might be some stuff in there e.g. uncommitted that you don't want used by 3rd party.
- Would be good for JAO (and NRAO) to have an expert opinion for both our AI apps and for our cloud tools like Slack, Outlook etc.
using AI to help with coding
- Modern LLMs for coding have made huge strides in the past few months. For simple use web UIs are fine, more advanced tools like claude can be very sophisticated. Certainly good for helping with e.g. scientist's matplotlib plots! Also good for making unit tests
AWS Innovate on March 13 - free online event to register for - have a look at the various "tracks" you can follow...

AOB - other AI/ML ideas/thoughts