Won't Someone Please Think Of The Astronomers ?

Yet another case of sweeping job losses to AI. This time it's the turn of the astronomers.

Who needs radio astronomers when you can have robo-astronomers ?

What am I talking about ? If you follow my content generally, you'll know I'm actually rather AI-positive. Indeed, this really fun political quiz scores me as more of a techbro than a neo-Luddite. 

Immediate disclaimer : this quiz is quite clearly deliberately a bit silly, so jokes notwithstanding, I have rather strong Views about being identified as a techbro. 

Still, I do maintain that LLMs are generally a net positive, so why am I claiming mass layoffs are just around the corner for us poor hapless astronomers ? Could it be the title was actually just clickbait ?

Yes, yes it could. Today's post is about "my" latest paper, in which "we" use machine learning to quantify HI deficiency slightly more accurately than previous techniques.

Well, that sentence has likely raised several questions. First, why the scare quotes ? That's because I'm last author on this work and wasn't involved with the initial investigation. I got added only after sending admittedly rather extensive comments and consultation on the observational side of things, so readers curious about the whole machine-learning aspect are advised to asking something or someone more knowledgeable about such matters. ChatGPT is probably a suitable place to start, being a machine itself after all.

But you might also be wondering about what HI deficiency is, why it's important, and why a small improvement is worth publishing. There at least I can be of help.

1) What is HI deficiency ?

It's actually quite a simple concept. Galaxies are found in a wide range of environments, but we often simplify this into just two categories : great big clusters, and everywhere else. Rich clusters are crazy, chaotic places, where you might get a thousand or more galaxies crammed together in an orgy of destruction and wanton stellar violence. "Everywhere else" consists of smaller groups, typically a few or maybe a few tens of galaxies, and sometimes just one or two. The violence there tends to be much more sedate and drawn-out. 

We tend to refer to everything-that's-not-a-cluster as the field. This includes individual isolated galaxies and small to medium groups, but when you get to agglomerations of more than, say, a hundred or so, you're into cluster territory. The numbers are very loosey-goosey, but there is a real difference between clusters and groups. In clusters, galaxies are moving much, much faster relative to each other, and the clusters themselves tend to have their own diffuse gas. 

Not shown here are the speeds. In small to medium groups these might be ~200-400 km/s or so. In large clusters they can be in excess of 1,000 km/s.

The diffuse gas means that a cluster isn't just a big group. A galaxy slamming through this hot intracluster medium experiences ram pressure, which is the means by which big clusters can be incredibly effective at stripping even massive galaxies of their gas, sometimes with just a single pass through the cluster. In groups, there might be some level of diffuse gas as well, but it's much lower density and the galaxies there are all moving more slowly. This makes ram pressure much weaker and less important.

Not that groups are entirely safe spaces for galaxies though, as that would be far too woke. Galaxies in groups are more vulnerable to tidal interactions which can sometimes be much more damaging than ram pressure. But broadly, as a zeroth-order approximation, galaxies in groups don't experience gas loss while those in clusters do.

HI deficiency is simply an attempt to quantify how much gas a galaxy has lost. Through large, carefully measured calibration samples of galaxies in the field, people have found relations between the size and shape of galaxies in relative isolation and their total amount of HI gas. This means that when we find a galaxy in the cluster, once we measure its size and shape we can predict how much gas it would have if it was living happily in the field like a sort of cosmic cow. The difference between the amount of gas we expect and the amount it actually has is the HI deficiency. I wrote a short calculator which gives a bit more details on this here.

Not that kind of field ! Though I suppose galaxies in the field do eat gas rather than grass, so maybe the metaphor is more punny than I intended... 

For the enthusiasts, HI deficiency is a logarithmic parameter. A galaxy with a deficiency of 1.0 would only have about 10% of the gas content of a field galaxy, if it had a deficiency of 2.0 it would have just 1%, and if it had a deficiency of 3 it would almost certainly be undetectable because it would have hardly any gas at all.

2) Why do we want to measure this ?

Paycheques, mostly, but also genuine interest. You can't form stars without gas, so the gas content of a galaxy affects potentially everything that happens to it. And knowing how much gas a galaxy has lost means we can understand not just what it's going to do next, but also how it's already been affected by its environment.

I mentioned that galaxy groups weren't safe places to hang out, but their exact role on how galaxies evolve is still rather controversial. Our basic picture is that galaxies are largely born in the field and clusters assemble gradually as galaxies fall in together, whence they're likely to loose all of their gas very quickly. Clusters, be their orgiastic nature, are the sexy bits of galaxy evolution, but just like real orgies, they're highly atypical of normal behaviour*. So studying galaxies in these weird environments is a bit like going to a brothel and asking the participants about their favourite movie to watch with their other half when they're having a lovely night in.

* Only a few percent (low single figures) of galaxies live in clusters.

If you do want to study a galaxy having a lovely night in, that is, a normal, typical galaxy, you then need to go to the field. There it appears that galaxies are far less susceptible to gas loss, but there may be some level of pre-processing even so : they can still lose gas through tidal encounters, just at a much lower rate than through ram pressure. Whether this can be at all significant, i.e. whether it can be enough to affect a galaxy's star formation activity prior to reaching a cluster... that's the bit we don't understand.

In fact it gets a lot more complicated than that, for (at least !) two reasons. First, when it comes to galaxies, size really matters. A massive galaxy can be far more resilient to gas loss than a smaller one, which might even be able to expel its gas through its own stellar winds and supernovae explosions.  

Second, the star formation activity of most field galaxies appears suspiciously stable, as though they were being continuously resupplied with gas to compensate for any losses due to star formation. This is known as accretion, and is widely accepted to be a thing that happens, but proving any specific examples of it – that is, actually spotting gas inflow while it's occurring – is bloody difficult.

And third (I said "at least" two and I meant it literally), the outermost gas of galaxies can be a lot less dense and easier to remove than the stuff within the stellar disc. So this may be vulnerable to ram pressure and other gas removal mechanisms even in relatively low density environments, a prospect known as starvation (or strangulation). Whereas our classical sort of ram pressure shuts down star formation almost immediately by removing all of the gas in a galaxy, starvation just depletes its reservoir, not the gas that's actually currently forming stars. It's a bit like draining the fuel tank versus removing all the fuel that's already actually in the engine.

Potentially, a galaxy might experience all three effects. It might grow initially if it's in a low-density part of the cosmic web, then as it accelerates towards denser regions it might lose its outermost gas. Only as it falls into dense clusters will it experience the full power of ram pressure stripping, though all of this depends on the mass of the galaxy as well as its environment.

Measuring deficiency as accurately as possible is therefore important in helping us understand all of this. If we could measure it precisely enough, we'd have a much better idea of the effects of environment on galaxy evolution.

3) How do we go about measuring it ?

It's not that difficult. The equations that people have come up with typically involve just two parameters : the size and shape of a galaxy. That's it. From these, we can calculate how much gas the equivalent isolated galaxy would have, compare with our target, et voilà, we have ourselves a deficiency.

The real-world problems are where it gets tricky. Until recently, detecting HI wasn't that easy, and even now, the total number of detections is probably no more than 100,000 : statistically large, but vastly smaller than the number of optically-catalogued galaxies. Parametrising the environment, so that we can really determine which galaxies are the helpful hermits we need for calibrating the relation, is also far from straightforward. And measuring size is relatively easy, especially given recent automated advances, but shape remains something that's still largely done by eye.

This means that of the half-a-dozen or so brave attempts to produce the relations we need for measuring deficiency, every single one of the buggers has come up with quite different values. And all agree that we really can't get any more precise or accurate than this. Basically, all this work has meant we can come up with four broad categories :

• Negatively deficient galaxies, which have more gas than expected (these are rare)
• Non-deficient galaxies, which we hope are typical
• Moderately deficient galaxies, which definitely do seem to have lost some gas
• Strongly deficient galaxies, which have lost a lot of gas

So I lied. "It's not that difficult" is only true in the sense that when you've got someone else's relation, it's easy to apply this to your own data. But actually establishing those relations is something that involves a lot of careful work. It can't be rattled off in an afternoon, there's no clear reason to prefer one person's relation to another except for blatant favouritism, and all that work doesn't get you a terribly informative result anyway. 

Perhaps, though, there's a less biased approach that might also increase the precision of the measurements, at least a little bit.

4) Can robots help ?

That's where machine learning comes in. Morphology is hard to measure because you need precision estimates of things like how tightly the spiral arms are wound, the arm/inter-arm contrast, the bulge to disc ratio, etc. etc. You can do all this very easily by eye, but automation is fiendish.

What if we just chucked a bunch of data at it ? You know, the old approach of throwing everything against a wall and see what sticks ?

This is a time-honoured plotting technique when you haven't got a clue what's actually going on.

That's essentially what we've done here. Rather than trying to measure all the complex parameters that define morphology – something that nobody ever seems to have done properly – instead the lead authors collected a bunch of parameters that, being relatively easy to measure, were already available for a large sample. Things like the brightness in different bands, estimate of the concentration, and all combinations of those parameters. Whereas a human would struggle to look for trends in more than a few independent parameters, and collapse in a screaming wreck if they had to try all the different possible combinations, for machine learnings, this is apparently no problem at all.

I'm not going to pretend I understand much about machine learning, though I think the description in the paper (section 3.4) is pretty good. The hope is, though, that there are sufficient measurements here that they effectively encode all the morphology information without having to make it explicit*. To determine their isolation we used someone else's big, popular catalogue, which after some fairly detailed checking does appear to give values which are in good agreement with other definitions of "isolated". And for the HI we had the giant ALFALFA catalogue, which rather surprisingly hasn't been used in his way before.

* I did suggest to include the actual morphology as conventionally determined, which we actually do have for the sample, but this was never implemented as far as I know.

The definition of isolation is actually quite tricky. Basically you want a galaxy which is sufficiently isolated as to have likely experienced no previous interactions, but determining this is not so easy. For this, you need 3D information about not just a target galaxy, but all of its surrounding neighbours : galaxies which look close together on the sky might actually be far apart along our line of sight, and vice-versa. Unfortunately getting this information for every galaxy is just not going to happen (though maybe some next generation surveys might help), so this is certainly one of the main limitations.

One of the other more recent efforts came by a very different approach in 2018. There the authors were much, much more careful to define their isolated control sample very carefully. Even thought they actually had less galaxies than some previous studies, they showed that they were able to decrease the scatter in the final HI deficiency estimates.

Our approach, which is nicely complementary, is more to go in waving our arms and screaming "let's throw MOOOAAAR DAAATTAAAAA at it !!!". Thanks to other people's catalogues, we have something like a order of magnitude larger sample size of nearly 7,000 isolated, gas-detected galaxies. The hope is that this will balance out any imperfections of likely including a few not-truly-isolated galaxies.

5) Does it work ?

Yeah, it seems so. Quite well, actually : the scatter reduces, not by a massive amount to be sure, but it does decrease.

Figure 6 from the paper. This plots the deficiency of isolated galaxies, which should be roughly a Gaussian centred on zero. The "RF" model is the main result of the paper, which does a bit better than the "linear model" (our own, using the more traditional two-parameter approach) and that of Jones. A small improvement is still better than no improvement !

Actually here the referee had a really interesting point, one that had not occurred to me at all : how do you know that the reduced scatter is real and not a statistical artifact ? Specifically, if the model is predicting a narrow range for the expected mass in the galaxies, it will reduce scatter artificially – just because a smaller predicted range means less outliers.

This one took me a good long while to get my head around but it appears to be a sort of "regression to the mean" effect. The essence of it is that it might be reducing scatter by reducing the predicted range, not by actually being any more accurate. Forget the deficiency bit for a moment and let's consider the isolated galaxies : the ones where we expect no gas loss, which we use as our control sample.

That is, suppose our isolated galaxies have HI masses of 3, 4, and 5 in whatever units (simple numbers to keep things simple). The ideal case would be a model that take the optical parameters of each galaxy and predicts corresponding masses of 3, 4, and 5, which would give deficiencies of zero in each case. A model which predicts, say, 17, 69, and 78 would clearly be no good and have a huge scatter.

But a model which happens to give the average mass of 4 for all objects will be much better and reduce scatter... even if it's only doing so by averaging. It might not really have any true individual predictive power, it's just that for the whole data set, the scatter goes down.

Fortunately it seems that this is not a statistical artifact. When we apply the model to isolated galaxies not in the training data, we find much the same range of predicted gas content. Since those objects are independent of the training, it appears that the model is indeed actually predictive.

And when we apply it to the non-isolated galaxies, we find all the classical effects that HI deficiency increases with galaxy density. This is a good sanity check : we haven't uncovered anything new, but we demonstrate that the model gives sensible results. Galaxies in richer environments have generally lost more gas. Hooray !

6) What's next ?

Well, we could try throwing more data and more parameters at it, but it's not clear how much this would help. Reducing the scatter further is probably possible, but how much is very hard to say. We might even be starting to hit a limit.

The problem is that as galaxies lose gas, they evolve in response. If they just lose a little gas, they'll get a bit redder as they stop forming hot, blue, short-lived stars at such a high rate. They won't turn fully red, but they'll redden and their size is also likely to change as their outermost gas gets removed. 

How quickly these changes will proceed is uncertain. And that's a problem. Suppose we measure a galaxy just after it's lost half its gas. If the gas loss has been sudden and recent, it will still have almost exactly the same colour and size as when it had all its original gas content, so we'll be able to measure it as being moderately HI deficient. But if we wait, say, a billion years, then its colour and size will definitely change. At that point our HI deficiency measurement might be quite different, because the model is predicated on the idea that we detect galaxies soon after the stripping occurred. The problem is that the calibration parameters themselves change due to the very effect we're trying to measure.

To be fair, in rich clusters it's likely a safe assumption that gas loss was sudden and recent. In those environments gas loss is so strong that if you detect an object which is deficient, it's quite likely either lost its gas recently or is still in the process of doing so. But even in clusters, galaxies are on orbits which take them well away from the region of gas loss for a time, and in lower density environments, this assumption may well break down completely.

Now to people who work in this area, this may be "intuitively obvious", as the referee stated. But honestly, I work in this field, and it wasn't intuitive to me at all. I'm quite pleased that we managed to include a qualitative estimate of the size of this effect in the paper, because I've never seen it described explicitly anywhere else besides a few (extremely) vague allusions here and there. It definitely seems like this will indeed contribute to the intrinsic scatter, though whether this will put a hard limit on the precision of what we can do is difficult to determine.

---

Unfortunately, what we haven't been able to do is show if we can use this to find anything beyond what was already known. For example, do we find any objects of strong deficiency outside clusters ? Do we see moderate deficiency levels in groups and filaments, i.e. pre-processing ? This was partly due to this being a lot more work – and the lead author has some strict time constraints – and partly because the reduction in scatter is nice but it's not massive.

What we actually do next is not so easy to judge. There are improvements to the modelling techniques which might help, and we might add the morphology parameter directly, or use larger data sets from FAST. All this is more dependent on our own availability (read : willpower) than anything else.

Thankfully, though, we have one clear conclusion. For this particular brand of well-trained robots, the risk of massive job losses and societal collapse appears to be acceptably minimal. We can all rest easy in our beds tonight.

Killer robots yearn to steal the glory from radio astronomers for slightly reducing the scatter in HI deficiency measurements. Look, James Cameron, this would make for a better movie than Avatar, for crying out loud.