Hello, everyone. Thank you for coming out on a nearly sunny afternoon to sit inside this lecture theatre for a very special electron. Chris FLINTOFF I'm a professor of astrophysics here at the department, and I'm one of the people organising what's become a wonderful series of waiting workshops sponsored by Philip Whetton, who's is here. And for him you support we're very grateful. And this workshop is on the topic of planning for surprises.
So it's an attempt to bring astronomers from different fields in different places for parts of the universe, I suppose together to talk about how in this world where we have access to huge data sets, we can still be surprised. We can find some things that we're not yet looking for.
And so today, those of us at the scientific conference had talks about distant galaxies, about cosmology, the science of studying the universe as a whole, nearby galaxies, galaxies where you can distinguish what shape they are. We had talks about radio astronomy and we had kicked off with talks about acceptance planets around other stars. And that's where we're going to spend most of the next hour, an hour speaker who's going to talk on the topic of how to find planets around other stars.
And delighted is David Hogg, who joins us. Fresh off the plane from New York, David started his career at MIT. He was at Princeton at the Institute of Advanced Study, told it's important for my dad that I mentioned that I never spent the rest of his time in New York, though the website of New York University says that he spends most of each summer in Heidelberg.
And most of it I'm sorry, some of it somewhere in Heidelberg at some of each week at something called the Flatiron Institute, which is a new place in New York that's got the astronomers together to think about problems of big data. So there really is no one better to talk to us about the challenges ahead. And I invite you to find out exactly how we do find planets around other stars and Uranus. Thank you, Chris. That was very nice. Good. Let's get into it, shall we?
So I think you probably heard a lot about machine learning and you have to be living under a rock to not have heard about machine learning. But let me just give it a little let's say a few words about what machine learning is, and then I'm really going to criticise it, but we'll talk about anyway, we'll see where it goes.
So say you want to find all the kittens in all the YouTube videos and this sounds like a joke, but actually one of the first really big successes of deep learning was a demonstration by Google that they can find all the kittens in all the YouTube videos. And the way they did it is they essentially they they built they took this thing, which is called Deep Network or whatever, but it's essentially an extremely flexible function.
And they found of the parameters of an extremely flexible function that can take a video in is out input and return a boolean. Yes, there's kittens or no there aren't kittens and that is true. Now, if you think about how complex videos are and how different they are, that is no mean feat. It's remarkable that they did that, although you can wonder why.
But then fundamentally the reason it was possible is that they owned YouTube and so they could use any enormous data set of videos both with and without kittens, and to train this highly flexible function. It was the enormity of the data that made this possible, and that has a lot of connections to things that Chris and I talk about. We talk a lot about open science and open data, and we work a lot on kind of making astronomy more open.
And one of the problems with these kinds of things is it really depends on what data you own. And notice the big leaders in deep learning are companies that do not share their data. And so it's an interesting question. What does this kind of model or this kind of operation have to do with science? And I think in its basic form, in that basic form, I think I would say not that much. Now, some people at the workshop here at this week will disagree with me strongly.
And so people at the workshop are working hard to exploit these methods for science. I'm going to give a bit of a vision about that at some point later in the talk. Who knows, by the way, who here has an undergraduate degree in something relatively technical? Okay. I am going to disappoint you. Good. So I am going to talk about planets around other stars. I'm going to call them exoplanets in the business.
We almost all call them exoplanets. But whenever you hear X or Y, that's jargon, but whenever you hear it, just think planets around other stars and just if you don't know about them, they were first discovered in the nineties. You probably I mean, the only thing that's been in the newspaper as much as deep learning is exoplanets. So basically and I'm going to make some fun of that later.
And the first was out in the nineties and there are now thousands known which, you know, many people in this room are very young, but that's a pretty rapid advance. It went from being a very niche thing that nobody thought would work in the early nineties. In fact, there are talks. I remember going to seminars saying there's no way we'll ever detect mind around another star. 215 years afterwards, two thirds of the astronomical community was working on exoplanets for at least part of that time.
It really, really changed our community. And the reason it changed our community is that it is completely new class of objects. For one, we're always excited about new technologies. Two amazingly rich observationally, and I'm only going to give you the tiny sliver of that, how rich it is. Observationally and three, we live on a planet which makes it seem really important. You know, most of the things we do really have no applications and this is not going to be about applications here.
In fact, that that kitten slide was the last application slide we have. We now know there are billions of planets in our galaxy. We basically know there are more planets then stars now. I mean, it's a little debateable exactly what you would say for the populations, but we but correcting for the selection of X and so on, we now believe that there are more planets and stars and and there are some regimes of planets, some kinds of planets we are not yet sensitive to.
So I think once we become sensitive to those, it will be many more planets and stars. Do you agree with that? Yeah. And you could object to that. Okay, good. How could we know about these planets? Now, before I get there, I'm going to just remind you. Just remind you, in case you've forgotten any astronomy you learn. First of all, we orbit the sun. We're on a rock. Actually, we're really on a hunk of metal, but it's metal and rock. And we orbit the sun and we orbit the sun once every year.
Depending a little bit on your reference, friend. I won't tell you that. In fact, it doesn't. There's no truth to whether or not there is goes around the sun and the sun goes around the earth. But we'll come to that later. They will do that in question period, if you like. And the sun is millions of times larger than the earth. The earth is tiny relative to the sun.
And that's going to be an important part of the story, because finding planets is hard wise, finding planets hard because planets are tiny. And so we think of the earth as big. But you can fly around the earth. You can't fly around the sun for more reasons than once. The sun is orbiting the Milky Way. The Milky Way contains billions of stars, tens of billions of stars.
And it depends a little bit what your definition of star is and so on. But many billions of stars in the Milky Way, the solar system is about 5 billion years old and we know the Earth is 4.6 very accurately and the universe is 13, 14 billion years old. So it's interesting thing that the earth is a significant fraction of the age of the entire universe. Just things to remember. Keep in mind, they're just some context for thinking about what we're working on.
Okay, good. So in our solar system, the rocky planets are on the inside and the gas giants are on the outside. And all of my early scientific life, we believe that must be the way it is, because after all, the solar system is that way and everything seems typical about our star. So probably they're all like that. It's that turns out to be totally wrong. The inner planets get their heat from the sun.
We are heated by the sun, although there is residual heat in the inside of the earth, as Icelanders know. And. But the outer planets get some heat from the sun and some heat from their original gravitational collapse. Jupiter is still cooling down from its original class, and there appears to be a continuum between planets and stars.
There's no real distinction between planets and stars in the sense that the lowest mass star like objects we can see look very much like the highest mass planets we see. Looks like there's just a continuum between planets and stars, which is strange and was very surprising to the community. Good. How do we find plants? There's three big ways and then many little ways.
Actually, there's about to be four big ways. But right now there's three. The one that happened in 1995 is radial velocity measurements. And the idea is the planet. And this the star and the planet orbit a common centre of mass. And so when when the planet is accelerating around the star, the star is also accelerating around the planet. Now, the star's accelerations in the stars. Velocities are very tiny, but they're not zero.
And so, for instance, the sun moves something like ten or 15 centimetres is second in response to the planets in the solar system. So if you can measure the sun really accurately and we can of course measure our own sun that accurately, but that's not so interesting.
And then radial velocity is you can measure velocities very accurately in the world because of the Doppler shift that things are blue shift and the red shift and blue shift when they're coming towards the register and then going away from you. And it's possible to measure those red shifts and blue shifts very precisely. So that was actually the first method that found planets around other stars. And there have been. Oops. Good. And there have been hundreds of discoveries from radial velocity.
And many more than hundreds, thousands of stars have been studied carefully with radial velocity measurements. The big game is transits. If we're very lucky and the planet is orbiting the star in such a way that it passes between us and the star, then it blots out a little bit of the light each time it passes in front of the start star and blots out a little bit of the light. And those periodic transit, those are called transits, those little eclipses, they're like mini eclipses.
And those eclipses are periodic signals that are imprinted on top of the stars brightness. So you can measure the brightness of a star unbelievably accurately. You can see these little blip as the planet goes in front, and an earth like planet blots out 100 parts per million, meaning a part in ten of the four 1/10000 of the light of the sun.
So if some other astronomer on another star is looking back at us and sees the Earth Transit, it would block out ten to the minus four of the light of the sun for 13 hours every 365.25 days. So if the other astronomers have found us, they were very persistent. But the crazy thing is that NASA's Kepler spacecraft has found thousands of planets this way. And there are other missions both on the ground and in space that have also found many.
So this has been the most productive, even though it requires this amazing coincidence that the orbit of the planet lie in along our line of sight. So it's only a tiny fraction of planets we can see this way. It's still been very productive, and I'll try to give you a sense of why. Of course, what we really want to do is just see the damn planet. We don't want to see it indirectly through the wobble or the light it blocks.
We just want to see it. And so this is one of the holy grails for astronomy. And there have been a couple dozen planets found just by being directly imaged. You can just see them. But unfortunately, all the ones we can just see right now are very young planets. Planets that are so young, they're still very hot like the nebula they formed in. So we're really seeing very special planets when we directly image.
Now it is on the roadmap for NASA and ESA to directly image that much, much smaller and much more normal planet. So. So this dream is alive. And there's a lot of astronomers right now in the United States working on mission concepts for this, and there's many other methods for finding them. A bunch of plants have been found through gravitational lensing, which is an absolutely wonderful thing, which are not to talk about at all.
There's some planets have been found. A bunch of planets have been found because young stars have accretion, disk or gas disks. The planets form in gas disks around the star like our own solar system. Why is it in a plane? Probably because it formed out of a disk of dust and gas. And you can see in some young stars those disks of dust and gas. And you can see they're perturbed by planets. There have been planets found by their dynamical perturbations on other planets.
And that's actually becoming a more and more productive way of finding planets over time or finding new in more and more ways to find planets through dynamical perturbations that go beyond the radial velocity methods. There have been planets found by pulsar timing and pulsation timing. And the really big thing is astrometry. They the you may have heard that the ESA Gaia mission just released very, very detailed information on 1.7 billion stars. This is just an early data release.
When they do their late data releases, they will detect planets through the astrometric wobble of the star, just like the radial velocity wobble. There's also a wobble on the plane of the sky, and it's believed that Gaia might find tens of thousands of planets and might be the most productive producer of planets. So we don't know yet. Actually, it's one of the things I'm working on.
Gaia, by the way, the guy database is very exciting thing, a very big moment for open science, a very big moment for astronomy. If you're interested, if you're an astronomy buff, look up the news on the guy I mentioned. It's really been remarkable and it's all the data. I've only been out for seven weeks and it's already had a huge impact on what we believe about the Milky Way and about stars. Good. I said this. Water exoplanets, they're planets around stars.
But from my perspective, they are exceedingly tiny signals in exceedingly boring data. And I'm going to try and say why that work? Why does the data we kind of want the data to be boring, it turns out, because the exciting discoveries are easier to find in the boring data. What aren't exoplanets? They're not of this world and they're not in this. These are all massive press releases. But check this. They're not too bad. That is not what exit plans are.
I mean, maybe for somebody they are actually, it was I was I went to the I don't know if people anybody here went to the eclipse this past summer. I was in central Oregon and a campsite. And then after that, we went to the Oregon coast. And I was with one of my colleagues who's in computer science who works on exoplanet detection with me for standing on the beach. And there's this mossy cliff with water dripping down it and there's whales reaching out in the ocean and the sun is starting to set.
And he said, we have a little bit of hubris to think we know what these things are that we're detecting in the data, just looking at the richness that we see on Earth. But that is not what we've discovered. Okay, good. So I'm going to talk about Kepler mission because I'm going to focus on transits. But almost everything I say about transits will apply very much to all the other methods we've applied. Everything I say here, we have also applied to direct detection and everything I say here.
We've also applied what we are applying to radial velocity and everything I say here, we're going to apply to Astrometry. So all of the different methods have the same kind of issue, the same kind of issue when it comes to data analysis. And I'm an I'm a data analyst, by the way. I'm like a software person. My my group is a software group and a methods group.
We do a lot of statistics and we do a lot of data analysis and we build a lot of code and we all do open source, everything's open source and everything that I I'll show a few results from it, but not much. But the results I show will be results where you could get clone it from GitHub and type go and you will get those results. Good. The Kepler mission was unbelievably simple. That's what I said about boring.
It just stared at 150,000 stars and delivered a brightness measurement every 30 minutes. That's all it did. Okay. It did a few other things, but not much else. I'm looking at Deirdre over there because he's one of the people of the Kepler mission office, and it lasted for 4.1 years. In its main mission, it made something like 10 billion measurements. During that time, it found thousands of planets, as I mentioned.
And it's been followed by the K2 mission, which was a repurposing of Kepler after it basically got damaged. And that's how it ended its 4.1 year mission. But it's lived on. And actually the interestingly, the commission has been even more productive with some damaged satellites, which is another thing that we think about a lot.
The group that's gathered here for this week for the wedding workshop is one of the themes of it is kind of repurposing things, repurposing data and repurposing hardware to do new things. And so that's a beautiful repurpose. The K2 mission actually has some responsibility for it. And here's how. One of the reasons that Kepler worked so well is that it's an unbelievably simple object. It's a telescope. So now, you know, it's a laser. So this is a telescope here.
And then it's got its solar panels over here and it's got the only moving parts it has. I think this what I'm about to say may be slightly false, but essentially the only moving parts it has is for reaction wheels. And those were the things that failed to end its mission. Two of them failed and its primary mission. But it's an extremely simple device and it's just a telescope, very simple telescope, going to a very simple focal plane, that focal plain and simple.
But it was very expensive. But the mission overall, because it's so simple, was very cheap by by like space launch standards. This is not an expensive mission. And in its main mission, it just stared at this one part of the sky. Now, the professional astronomers in the room will have no idea where this is on the sky, but some people in the audience might. Now, I certainly doubt and this this plot is way out of date. Anybody who works on Kepler is shaking their head that I'm showing this plot.
But this plot is from the paper that I was a co-author on. And these the blue and black points are essentially all detected planets. The difference between blue and black has a little bit to do with the certainty with which they've been designed as planets. But but the story that's emerging is basically everything blue and black here is a planet, and I'm showing it an orbital period in days.
So one year is like here ish and I'm showing you an orbital radius here where one is an earth radius and there's an orange dot at 365.2, five and one because the orange dots are the planets in our solar system. Remember, there's eight planets Mercury, Venus, Earth, Mars, and they go to Jupiter. Right. And all of these and know notice so many interesting things here. Look at these planets and Kepler found thousands of planets. Look at this. There's Mercury, the fastest planet in our solar system.
Notice that almost all the planets that Kepler detected are faster than mercury. And many planetary systems have multiple planets inside the radius of mercury. So it looks a little weird. Another thing, another thing you can notice here is here's Earth. The goal of Kepler was to find planets that are like Earth. And notice that. What are these lines? These lines are kind of hardness. These planets are easy.
Well, these are really easy to find. These are a little harder. These are a little harder. These are a lot harder. And so the reason that the planet population is dropping as we go this way is because it gets harder and harder to see them. Why does it get harder? Because as you go down, the planets get smaller and you go out and get fewer eclipses in 4.1 years. Right. You see. So you get smaller eclipses and fewer eclipses as you go out.
And that's why that that basically it's very hard to find planets around here and it's very hard to find planets in the vicinity of Earth and Venus and not even close to Mars. It's so good. That's just some context. I think we'll come back to that. I'll come back to these results at some point later.
This is actually key to data, not Kepler data, but it makes the point that I wanted to make, which is that the the the I'm going to talk about braces in parts per million and I'm going to centre them on zero. So think of this the following way. Imagine I was measuring a star and it had perfectly constant brightness. It would be set at zero. So this is like fluctuations in the star away from its mean behaviour.
Okay. So zero would be no fluctuations and then I'm going to write things in parts per million. Right. So so if you just take raw data off the K2, this is a K2. I like her, not a candidate like her. But if you just take raw data off of it, there's kind of many thousands of parts per million variation. But if you look carefully at this variation, it's highly structured and very structured.
That's structured because the spacecraft has configuration issues and the spacecraft is pointing and temperature is changing and that's projecting on to the data. And so one of the things we do is we kind of get rid of that spacecraft motion. And here's the saying light curve. This is the same star where we've kind of modelled. This is subtracting our best fit model of what the spacecraft is doing.
And you see now we're getting more like 100 parts per million. And remember, that's the level at which we're looking for our transits. So this is kind of just an illustration that there's data. You don't just pull the data off the telescope and find planets, and that's going to be part of the story. And then this shows like a periodic, you know, you might have to be an astronomer to see this.
This might look just like a noise to you. But if you're an astronomer, you see these little dangling little spikes. There's a data gap here that one of them and one of those down going spikes. So if you take the data and you fold it on that period, you see a transit, a planetary transit. So that's a planet that we found in the K2 data. Good. What did we learn from Kepler? We learned a huge amount from Kepler.
Here's just some highlights. These are very personal highlights. Different people working in the mission would give you very different highlights. And I have to say, I'm not in the mission. I'm an outsider. I write software and I use other people's data. One thing we learned, as I said, there are comparable numbers of planets of stars, maybe more, I think probably more is my current rate of the situation.
Many stars have very different planetary systems from our own, including these very close packed planets, planets on very short orbits, but also planets of around twice earths radius are the most common planets by far we now know. And there's no planet of that size in our solar system. So they're like they're like the factors of a few or maybe even ten more probable than Earth's and Neptune.
So we have an Earth and a Neptune, but we have nothing in between. And we and Jupiter, like planets, however, are very common. We're learning and it looks like a very large fraction of stars have a kind of outer gas giant and might be essentially all stars where the jury's out. But it's very basically every star we've looked really hard at for an outer Jupiter. It looks like they might have. Good. I said that. Okay, good. I mentioned this earlier.
If you're trying to find Earth, you need to find something that does 100 parts per million for 13 hours every 365, 3 to 5 days. Now, the problem is spacecraft variability, as I just showed you, is bigger than that. And also for many of the stars we study, including the sun, just the natural brightness, variations of the sun are larger than that. Actually, it depends a little bit on what phase the sun is. You're the sun.
Sun has a 26 year cycle and it has an active phase and then an inactive phase and it cycles back and forth between active and inactive on a 26 year period. Anyway, in its active phase, the sun varies by more than that, and in its inactive phase actually varies a little less than that. And actually one of the interesting things about this is it's easier to find planets around inactive stars, which is a long story that people are thinking that good.
So is it impossible to find Earths? Given that we're trying to find them in the face of noise, that is larger an answer to it. And obviously the answer is no because we found lots of planets that are getting very, very close to earth. So let's talk about why. The reason it's possible for us to find planets in the face of these noise sources is exactly the reason it's possible for Google to find kittens in YouTube videos. It is because we have enormous numbers of stars.
The first thing about Kepler was not that it stared for 4.1 years. The genius thing about Kepler was that it stared at 150,000 stars for 4.1 years. In fact, it would have been better if it had stared at a million stars. We would have done more and we would have done better because in fact, our inferences about the Kepler data are limited by the amount of data we have. You might think, well, 10 billion data points. Isn't that enough?
Well, it's not because we'd like to train flexible models, like the models that can take a video and tell you with a candidate, we want to take models that can take a star's light curve. Its behaviour over time and predict its future may take its past behaviour and predict its future, take its future behaviour and predict its past. We want informative models or informative functions that can predict the behaviours of stars. So there's sort of two aspects to this big data thing.
The first thing is we use the large amount of data to learn flexible models that can predict how a star varies. Stars vary statistically, but not unpredictably because they are convecting. It's a big ball of fire, it's got a convecting surface, and that surface has temperature and brightness variations that are stochastic, but they're not completely unpredictable.
And similarly similar but kind of orthogonal to the spacecraft is very and remember I showed you there is that spiky behaviour by the spacecraft. How do we figure that out? Well, we figured that out because if two stars very together, if I look at one star in the Kepler field, in another star in the Kepler field in general, these stars are hundreds of thousands of light years apart. They're not near each other in any sense. So they are not going to move in a synchronised way.
They're not going to do sinking. They're not going to synchronised swim across that whole field. So if stars vary in concert, we can use the covariance of the stars to learn about what the spacecraft is doing. But once again, you need an enormous number of stars because you have to see a lot of stars all move together to infer that that is coming from the spacecraft rather than from from of planet or from the stars.
So the idea behind the data science that we do in Kepler is we take the data, we use it to build a flexible predictive model of how stars work. And we use it to move to build a predictive model of what the spacecraft does. And that's how we find the planets. Now, that isn't quite the whole story. So we built these models. What how can a star very predictive model have started? Very we built the model. How is the spacecraft carrying now? We need a model for how the planet transits.
The nice thing is and the thing that's most important is our expectation about how a planet transits it is a very rigid expectation. We knew exactly what a transit should look like, have a very simple shape and they're periodic in nature. And so we are talking about stochastic models for the star in the spacecraft, and we're talking about a very rigid model for the planet.
And it's that mix of having a having a crazy flexible models under the stars, but a very simple, rigid, beautiful model for the transit that lets us find the transits in the noisy data and that. Problem structure is very general, by the way. That problem structure is also one of the things that is motivating the workshop we're doing at Christ Church because the it's easy to find the signals you're looking for,
it's hard to find the signals you're not looking for. And transits are signals we are looking for. So that's something we can talk about in the discussion. There's lots to say about that. It's very interesting. So it's really this contrast between the flexible models we use for the new sciences and the rigid model we use for the planets. That's what makes it possible for us to find the planets. And it is very related somehow to this question of where are the kittens in the two videos?
By the way, many problems in natural science have this structure. This is very common. Like if you work on, say you're trying to measure neurones, you're trying if you're doing real time imaging, calcium fluorescence imaging or whatever of of a slice of the brain and there's neurones firing. There's an immense amount of nuisance information, which is all the kind of noise in the image and all of the shapes of the neurones and stuff.
You're just trying to identify which neurone is, is, is, whatever they call it, spiking at certain times and you're trying. So there's a very simple thing which you're asking for certain very, very well defined spikes are trying to draw them out of the data, but the data contains tons of nuisances. So it's a very common structure for problems in science. And it is often the case, especially in physics, that the things we care about have a very simple form.
And so so my view is that we can harness machine learning to and to help us solve these problems because the machine learning can handle the part of the problem we don't care about. And if you go back to the kittens, why did I say that? I didn't think the kittens finding the kittens YouTube videos didn't seem like science to me. Why not? Because we didn't learn anything. We didn't learn anything about either kittens or YouTube videos. From that experience, we just learned that Google's awesome.
What we need is when we're science, when we're when we're when we're scientists, we're trying to understand where we want to come to some new understanding. We want to make new discoveries that fit into some other understanding. We understand things about the about planets by finding the planets. We don't care about the stellar variability. Actually, some people do care deeply about stellar variability. But in in this context, we don't care about stellar variability.
And and so we can use a model that leads to no understanding. It just handles it. By the way. So as I said, I was going to disappoint the people with technical backgrounds in the room. I did put a set of words on this slide. So if you want to do Wikipedia searching of things that's relevant. There's a if you go sort of the way we kind of think about these machine learning things.
Most of the work we do is actually with very, very old technology, which is pure linear models, and linear models have immense capacity. So if you're interested in going down the sort of how does machine learning work and what would be how do I understand machine learning? I would start with linear models because it's just incredible what general linear models can do. And then we go up to Gaussian processes, which are kind of a generalisation of linear models to very large function spaces.
They're very flexible things, they're like deep learning, but they have kind of better scientific properties in some ways. And then of course, there's deep learning, which is the thing you've always seen. And in, in the Star case, the, the, the, the most interesting technology is called recurrent, which is about imposing a certain kind of symmetry.
And then there's a new technique that's emerging in machine learning called generative adversarial networks, which is likely to have a big impact in these kinds of projects in the future, but early days. So if you're interested, there's some words to chase down. Okay, good. Everything I said about transits also applies to radial velocities. When people when we measure radial velocity is we're trying to measure the way of loss of a star too much better than one metre per second.
The Holy Grail is ten centimetres per second because it ten centimetres per second. You might be able to find Earth. And a star is a big, nasty object. The surface of the star, so the surface of the sun, the little patches of the surface of the sun are moving at a kilometre a second. The typical kind of convection speed is a kilometre second. We're trying to measure the surface of the sun to a metre per second, so we have to average over the sun and then or make a predictive model.
We have to average over the motions on the surface of the sun or make a predictive model. And so, of course what we're working on is trying to build a predictive model of stellar surface speeds. And then there's also some terrible things about the atmosphere that come in. The tiny details of the atmosphere have a big impact when you're trying to measure things. After all, the wind is a lot faster than one metre per second.
So, you know, that says that there are going to be imprints on the spectrum that are at the right level or at the bad level. But once again, elliptical orbit signature is very rigid. So we're competing a very flexible model to understand the stellar surface and the atmosphere. And we officially don't care about the stellar surface. We officially don't care about the atmosphere. All we care about is finding the planet. So it has the same structure and we're exploring the same place.
I mentioned Suzanne Abrams is right. There is one of the world's experts in thinking about those questions and. Good. I need to raise an epistemological point, and I want to end on this theological point, because I think it's something that's worthy of discussion. If you think about the most interesting planets that have come out from Kepler, many of them are the ones that are most similar to Earth, meaning the ones that are small and the ones that are on long periods and they're rocky.
By the way, we know those small planets are rocky now and through very interesting measurements, which I do have to ask answer questions about, the planets that are most interesting are in many cases the ones that are hardest to find. And if you ask yourself, well, how are we sure that those really are planets and not something in the data that's fooling us? Well, what's the most reliable things you can do? The most reliable thing you can do is go observe it with a new mission.
But then you'd have to launch another Kepler. And Kepler is about as sensitive as a telescope could ever be of its type. It's not obvious you could make a telescope that's ten times more sensitive than Kepler. So it's possible that these most interesting planets from Kepler will never be verified externally, which is a little strange.
And and we are pretty confident because we are very common, because then we can externally validate the planets that are a bit more massive and a bit shorter period and a bit bigger than we can verify using radial velocity measurements. We're confident that we're not being fooled at the bottom end, but I have to say we don't have external validation. And for the scientists, external validation is by far the most important thing.
And I want in general and as I put this slide back up, because I just wanted to remind you that many of these planets are extremely interesting and they're basically impossible to verify. There are some clever ways these blast points that are below this line have been verified through timing measurements. So there are some ways you can do it better. But in general, most of these things that are interesting here in near Earth are very, very hard or impossible to verify.
So I wanted to just generalise that a little bit. It's a very interesting thing about astrophysics. The the epistemological status of astrophysics is very strange. And I'm in a physics department and I consider myself a physicist. But many physicists are made somewhat uncomfortable by astrophysics for a particular reason, which is that the objects of our study are incredibly remote and we don't get to manipulate them. You can't decide to have two neutron stars collide.
You have to wait for it to happen and you have to hope you're looking in the right direction when it does happen. There's no chance of sample return and even the sun sample return would be almost an impossible mission. But it's a great idea. We should figure it out. We could only do radar inside our own solar system and even that out to the outer solar system. And we dependent very heavily on chance.
What happens to hit the telescope aperture when the shutter happens to be open and almost everything comes to us through photons actually. Of course, an important thing over the last two years is that we now have gravitational waves and we have neutrinos. We have a few other tracers, but almost all of our information comes from photons. So there's no way to do controlled experiments or externally verify many of the things we believe in astrophysics, and yet we know a huge amount.
And in another talk that I give, I talk about how do we know that there are black holes in the universe is expanding and there's dark matter because all of those things seem so crazy. How can you know those things? And yet we know those things with immense confidence and we know those things with immense confidence because there's many different lines of evidence that show us that that those are the natural explanations of what we're seeing.
But no ones like how the black hole into the High Bay to check it out. And we know that planets are common and we know that our solar system is not obviously typical of those things. We know very, very confidently, even though we can't do any, they're never going to go there. Sorry. Breakthrough Starshot. Whatever [INAUDIBLE] that thing is. Oh, I'm not supposed to say things like that. And it's Oxford. And again, this is my last slide. What do I want you to take home?
I want to take home a few things. Planets are plentiful around other stars. It is a it is not a it's not they're not it's not something like, oh, wow, this star has a plant. Oh, it's so cool that the sun has planets now. It's totally generic in the sun and planets and it looks like planets are as numerous as stars. And many of the planets we find are very, very different from the planets we're used to. And our picture of how solar systems form was just completely up ended.
In fact, there now when I was when I started in graduate school, there was a picture of how the solar system formed. Now there's no picture of how the solar system formed. We just lost it because it was just completely ruled out by the first plants that were discovered. And, you know, maybe that come somehow comes back to this question alone. But we also only have one universe anyway, and planets are mainly found indirectly.
We do not directly see them, despite what massive press releases might imply to you. And I think new data science technologies are critical. And the fact that the world is burgeoning with new astrophysics data and also new data science techniques coming from lots of different directions, bodes very well for the future of these things. Thank you.