Yeah, we had a detector with some ML stuff, it worked okay, but the main issue is there isn't much training data.

Overall the false positives are quite low (visually, which is hand wavey, of course). I've not really seen a big event that was not real. I ran it for a while 24/7 and while there are "scintillations" there aren't any circular waves. False negatives are a much bigger issue. Something like an IRBM or SRBM doesn't really show up, and those are much more common.

Fwiw, this uses a bandpass filter to look for the ripples, which filters out a ton of noise. Looking at ionospheric depletion would be good and probably more sensitive, but it requires accurate models of what the ionosphere ought to do. I've tried a few things there with mixed results.