Date:
Time: 1:00pm ET (11:00am MT)
Link to recording
Passcode:
Attendees
| Attendees | Regrets | Notified |
|---|---|---|
Mark Lacy |
Agenda/Notes
VLASS imaging
VLASS imaging needs aw-projection to get sufficient accuracy in continuum images that need w-term correction to meet the survey requirements (about half of them). This makes the major cycle step very time consuming. Employment of a GPU-based gridder will speed up the gridding, but it is still projected to take several years to process the survey data. Can AI/ML help speed things up?
We see three ways to improve the speed of imaging:
- Decrease the time for the major cycle (AI/ML probably not useful here, needs optimization of GPU/CPU usage).
- Speed up the minor cycles. Clean is currently used for the minor cycles, which are only a small fraction of the processing time, but speeding them would help a little. Algorithms like POLISH may be of use here.
- Decrease the number of major cycles. Deconvolution is in an iterative loop, with minor cycles modifying the model, then the major cycles reimaging with the current model subtracted. If the model can be made more accurate, then, perhaps, fewer of the expensive major cycles would be needed.
Some interesting papers of potential relevance: