Send the link. All right, this time for real. This is SETI Live, live from the SETI Institute in Mountain View, California. My name is Bill Diamond.
I'm the CEO of the SETI Institute and welcome to another amazing episode of our SETI Live series where you get a chance every week to meet our scientists, meet our educators, and others from the Institute and our colleagues and learn about and hear about and talk about what we do here at the Institute in terms of, broadly speaking, the search for life beyond Earth, but all the astrophysics, planetary science, and astrobiology that goes along with it.
And here today, we're going to be talking about a program that some of you may be familiar with that we've been doing now for eight years at the SETI Institute called the Frontier Development Lab or FDL. And FDL is an applied AI ML research accelerator for space and earth science started by NASA coming out of the office of the chief technologists back in 2016. and running since then under the direction of the Science Mission Directorate at NASA since that time.
So this is a program where we bring early career PhDs in machine learning and artificial intelligence together with early career PhDs in various science domains to tackle intractable problems in space science and exploration, earth science, et cetera. The FDL program over its history has covered a wide range of topics from heliophysics, astrobiology, exoplanets, planetary science, lunar exploration, and more, even astronaut health, space medicine, et cetera.
So it's been an amazing program, and we're here today joined by several people from NASA who've been part of this program this year, and one person sitting next to me, Megan, who's actually not only an alumnus of the program from, I think, third year, 2018, and is now at NASA headquarters and leading efforts in applied machine learning and AI at NASA headquarters. So we'll learn more about that.
But we have a couple of exciting teams this year doing both Mars science, trying to integrate various types of Mars data that is basically gathered by instruments on board the Curiosity rover and other resources. and also a project looking at lunar volatiles. You may not know that there's water on the moon, but there is, and it has also different characteristics that are changing during periods of night and day. So we're going to hear about that.
What I'd like to do first is actually have each of our panelists just say hello, introduce yourself, where you're from, and then I'm going to come back to you, Megan, to talk a little bit about the FDL program broadly before we go into the challenges. Take it away. So I'm Megan Ansell. I'm a program officer in the Science Mission Directorate at NASA headquarters between the Astrophysics Division and the Planetary Science Division. And I did FDL in 2018.
And I'm glad to be back hearing about these challenges. And I'm Janice Bishop. I'm here at the SETI Institute and also affiliated with NASA Ames nearby. And I do a lot of remote sensing work on Mars, the Moon, and Ceres. And this summer has been a big focus on the Moon and the volatiles there. Hi, I'm Victoria Dapouillon. I'm working at NASA Goddard Space Flight Center. I am a data scientist working on mass spectrometry data from several missions focusing on Mars, Titan, and ocean worlds.
And I'm very excited for this project, FDL, that the team Mars have been working on. I'm Eric Linus from the Planetary Environments Lab at Goddard Space Flight Center. And I work on the software for many of our missions, like sample analysis of Mars. We're working on Dragonfly mission and some other missions and trying to direct us our future missions, make sure we're ready for the artificial intelligence that we're going to need. Excellent. All right. So thank you for that.
Megan, maybe you can tell us a little bit more about FDL and the premise behind it, what NASA was hoping to achieve through that program and what you think has been achieved and what's next, particularly, you know, given your new focus on leading machine learning efforts within the science mission director.
Right. So, I mean, I think the thing that stands out about FDL to me, and this is through a personal experience because I participated in FDL in 2018, I think when you're a research scientist, so I was an astrophysicist at the time, I was a postdoc and early career scientist trying to figure out what did I want to do, what areas were new and exciting. And machine learning was one of them. I was a postdoc at UC Berkeley.
And, you know, I think learning a new skill like machine learning, you really need a summer sprint like FDL where you can sit down. Like I was so fortunate. I sat down next to two people who had PhDs in machine learning every single day and I got to bug them with all my questions. And the iteration time was so quick. And I was able to get a postdoc the next year at this place called the Flatiron Institute in machine learning. And that's me. Yeah, it was it was really amazing.
It was like it was something that I didn't even think could happen. Right.
um and it felt very like a natural easy thing to do after going through fdl so I think giving you know astrophysicists planetary scientists space scientists that skill fdl is very special in that way to be able to do that in such a short short time scale and a lot of the people that I did fdl with I'm still you know I will still bug them with questions it's a community it's a great community yeah yeah and and I you know the There were a few things that NASA was curious about when
they came to us with the idea for putting this program together. And the first one was, can applied can artificial intelligence and machine learning be applied to basic research topics of interest to NASA and help either accelerate understanding or make new discoveries or move things more quickly in terms of data analysis and so forth. In other words, is this a useful and helpful tool for the kind of research and basic science done at a place like NASA?
The second question was the extent to which interdisciplinary teams of computer scientists and domain scientists could work together in an intense sprint-like environment and get something done in a short timeframe? And then the third question was, can we bring private sector interests into it, private industry into it, and get them involved as a public-private partnership to support the program? And in so doing, of course, support maybe their own interests.
Among the more, I'd say, impactful partners in that context of public-private partnership has been Google Cloud. Google has, from the very beginning, provided GCP cloud compute credit so that the scientists doing work in the FDL program had more or less on-demand access to supercomputer resources and data storage assets that have been, I think, transformative from the program.
You talked about fast iteration, being able to turn and churn different approaches to data analysis, and the Google Cloud resources have been absolutely irreplaceable for that. Equally, we've also had access to NVIDIA GPU resources for advanced data and analytics. So between NVIDIA and Google Cloud providing hardware as partners to the program, It's been extraordinary. We've had also Lockheed Martin providing data sets, subject matter experts. We've had Intel involved.
So it really has, I think, for the most part, addressed in a positive way those three questions that NASA had at the very beginning. And I will add to that. So NVIDIA, I think they still do this, but they will provide a GPU to people who apply for it. And you sort of have to pinky promise that you will make a computer that can use it and apply it. So through FTL, I ended up getting a GPU and building my first computer. Oh, that's awesome.
And then using that to do research that eventually got me future positions in machine learning.
So it was very... was very opening and I'd be interested in everybody's take on this um but to what extent would you say you know so you went through this program and you kind of became a a machine learning expert or certainly a machine learning savvy person and um to what extent are you seeing that ai and ml as as a tool and technique are becoming more integrated into you know the the pure science pursuits like astrophysics, astrobiology, planetary and science, et cetera,
where it's kind of almost anticipated or expected that when you've got a PhD in these, you've been doing some machine learning. I mean, I think the thing that I've been really excited about over the last year or so is seeing how, you know, I think the focus has been a lot on integrating machine learning into more of like the data analysis side.
So when we have the data that comes off the spacecraft, but I think there's a lot of things, you know, especially with large language models now, where it can be very supportive of scientists as well and also integrated you know even before very early in the mission concept design like I can really see machine learning being integrated throughout the entire mission life cycle and I'm really excited about that I think I think that'll be something to look out for in the future things like
edge computing figuring out what to send back or not to send back yeah yeah I think it's really exciting yeah janice you want to talk about that briefly sure yeah I guess for our lunar project we've been working on understanding the hydration on the moon for a while, which is sort of weird. Like you had mentioned earlier that we think of the moon as this dry place.
And so back in 2009, it was discovered that there actually is some water on the moon, but it's been a slow process trying to get from, oh, there is some, and we see this water band to really understanding it. Even the last couple of years, I've had a NASA project working on this with a team, but it's still been slow because we're located around the world. And having the sprint this summer has been phenomenal.
So having two scientists and two machine learning experts together, we've made way much more progress than we have in the past three years. It's phenomenal. And what we've really learned so far is, is it time for that? Can I? No, let's hold that thought. We'll stick briefly to the, you know, just the applicability of these techniques first. Sorry, we won't talk about that. So what issues with the data set are, there's a global data set for the moon.
And then there's some, this is at 10 kilometers per pixel resolution. And then there's some spots with 100 meters per pixel resolution, but not very many. So for machine learning, you always need to have enough data for training it. So there are a lot of levels of the machine learning that we're using. But we're able to compare spectral features that are associated with the water with changes from time of day because we have data collected during different times of day, which is really helpful.
And then there are other data sets from previous missions, like elemental abundances and mineralogy that we can get from the specter. We have, just working on the abstracts this week, we have some new data, and everybody's really happy about how the results have been coming through just recently. That's great. Well, this is one of the things about scientific research is you can tell it's going well when the scientist is like, can we tell you what we found? Can we tell you what we learned?
It's like, oh, okay, yes. But yeah, that's wonderful. Victoria, what are your perspectives on this?
um so for the perspective of the use of machine learning and intelligence artificial intelligence for in my case planetary science I really see it in my research as a tool like machine learning as a tool to help the scientists in their analysis and this fdl challenge that I'm sure we'll talk more for the mars team has been really proven this right and what megan mentioned uh since I started working at goddard five years ago it was How can we implement that early at the early stages?
So right now we use data that was already collected, and that's amazing. But how can we think about future missions to not only enhance this mission, but to enable future missions going further away? So that's also something for astrophysics, HWO, that the effort you lead, Megan. But that's really what I'm fascinated about for machine learning in these applications. Wonderful. Eric. Yeah, our concern is I see us in the lab building the next generation of instruments for the planets.
We have big plans, dreams, going to Europa, all these places. The mass spectrometers that go there just make too much data. And I just see this train wreck coming They're going to have, oh, we have this great mass spectrometer, and we're going to go to Europa, and it's going to be great, and they're going to only send 1% of the data back as things stand now. So if we're not looking at this and we're not tackling this problem now, in 10 years or 15 years, it's going to be much worse.
And I love Megan's idea, too. Start at the beginning. These missions, the missions Europa, it's got to be integrated from the outside. Yes, integrated from the outside. Well, it's interesting because I've had more than one person from NASA say, never mind the issue of how much are we going to be able to send versus collect, but also just how much data do we have?
And I think NASA would be the first to say we've got more data that's never been looked at than has been analyzed from our various missions over decades. And that problem, to your point, is only going to get worse. And this is why I would say, yes, indeed, integrating this thinking from the very outset of mission planning is probably essential. But all right, well, let's move into the topics themselves, because we have a couple of people anxious to tell us about themselves, perhaps.
But so yeah, the first thing we'll talk about is water and the moon. So this is a wait, what moment? But not more, not anymore. I mean, we've known this now for... a couple of decades, but it has profound implications. It is a curious reality when we realize there's no atmosphere there. It's not big enough to have a strong gravitational field to hold an atmosphere and it's dry and it looks like it's barren. So how'd the water get there and what have we learned?
Yeah, I think most of this hydration signal probably comes from solar wind, from protons and part of the whole space weathering alteration. But there could be some water that's also embedded in the rocks. A lot of minerals have water molecules, single H2O molecules that are trapped in the molecular structure, the mineral structure. And so something that our team has discovered this week, we've had a hunch that there are two kinds of water, and now we have the data that proves that.
So it's one thing to think that your data says that, but when you actually can prove it with machine learning and with algorithms, then. And by two kinds, you mean what? So there's one kind of water that's trapped and it's there all the time. And there's another kind of water that's mobile during the day. And the day on the moon isn't like a day on Earth. The day on the moon is 28, 14 days, right? About 28 day cycle. Yeah, right.
Yeah. So But when we're looking at morning versus midday versus afternoon, we see a change in the water. So now our team has been mapping this change. So the changing water maps with one mineral. I'll let them tell you about that tomorrow. And the more permanent that sort of minimum boundary water is trapped with another mineral. And that's really exciting. So I've been hoping we'd be able to get towards understanding which compositions are associated with the hydration components.
And so that's really exciting. That's wonderful. Well, so speaking of tomorrow, since Janice mentioned it, so tomorrow, the two FDL teams that have been working so intensely this summer here at the Institute, will be presenting their findings at our readout event. We will be recording that. And so those sessions, because if you're interested, I think you won't want to miss them. They're really extraordinary. The work that gets done by these teams is truly amazing.
And yet it's delivered in a way that is really accessible. So you can understand the science, understand what's going on and how the AI and ML tools were leveraged to advance our understanding. So that'll be tomorrow, and we'll record those sessions, and those will be up on the FDL website, which is FDL4, the number 4, AI.com. So you'll find those presentations there after the event is completed. So let's go to Mars. We go from the moon to Mars. Lots of instrumentation on Mars.
We have rovers, we have orbiters, we're collecting all kinds of data from different kinds of instruments. And if I understand correctly, this year's team is really trying to look at how to integrate these disparate sources and types of data. Yes, I'll start and Eric will continue. But basically, we're trying to investigate how we could leverage commercial data instruments on Earth and flight instruments on Mars.
So because we don't have a lot of data from Mars, like limited amount of samples were collected with the Curiosity rover with the Mars sample laboratory, Mars science laboratory, we want to see if we could leverage commercial instruments that are on Earth from Johnson Space Center, from Goddard Space Center, and use transfer learning techniques to train a logarithm on these data to then be able to infer the prediction, like chemical identification on data from Mars.
So that was really the goal of this challenge, and we have amazing results, surprise tomorrow, but that's very promising. So if I understand correctly, then one of the problems you have is that the instruments on Mars are very specific. They are making very specific types of measurements in their own formats as well.
And as a result of limited data sets, you can't necessarily immediately apply AI and machine learning tools to those data sets because they haven't been trained and there isn't enough training data. So what you're trying to do is change that. by leveraging Earth-based systems. But they have different data formats, and that's part of the challenge? Yes, that was a big part of the challenge.
The data we collected from commercial instruments on Earth in our labs and the data collected from Mars were Similar enough, it's the same type of mass spectrometry data, but the format, the amount of scans we have, how long the experiment is, was completely different. So the data processing part was a big, big part of this challenge this summer.
And the team really experimented several ways to visualize the data, to represent the data, to cut it, like a lot of preprocessing part that really enabled machine learning to be applied. So they made the Mars data that was not machine learning ready format into a machine learning ready format. It's a huge, huge work for the team. Yeah. I might just want to add to that that we want to thank the SAM scientists, the sample analysis of Mars scientists from Curiosity.
Yeah. Hugely helped us to label that data. Okay. So one problem was just dealing with the format, so computer science problem. But the science problem was how do we label these? How do we say what this was and what that was? And so that was a lot of preparation that Victoria and I worked with the SAM scientists, and they were extremely generous with their time. Yeah, well, one of the nice things about FDL is how collaborative it is.
And there are four primary team members on each team, two data scientists, two domain scientists. But then we have mentors, advisors, subject matter experts, people outside who, as you say, provide enormous amounts of help. But hopefully it's because they're all motivated by the same thing that they want to like do what Jana says, like we learned something new here. So maybe Eric, you could give a very brief description for those who may not be familiar with it. What is mass spec data?
What kind of information does that provide in terms of, you know, Martian atmosphere, Martian soil composition? What kinds of information is being picked up from the mass specs? The mass spectrometer, you can think of it as the nose of the rover. It can smell the soil and tell you what it's made out of. It breaks it down. It can take a sample from the soil.
um like just a solid sample from the drill drills it takes a little bit of this powder it's just a dust uh and it can tell you what molecules and atoms that that is made out of it's both elemental and molecular yes it tells you basically the atomic mass of elements of that thing and we can use it for either the the powder we can take the drill and sample or we can just take in inhale some of the atmosphere of mars and do it you really can just either sniff or sample literally sniff for
example yeah when we when we sample we actually burn we actually vaporize the solid sample to turn it into a gas so it's as if you know it's like uh you throw it in the fire and then you smell the fire so what something smells like Now, there are different instruments on Curiosity Rover. Are you also trying to integrate or understand the relationship between these data sets as well through this effort?
we uh yeah we in our workshops working with the uh msl science team we included the chemin team yeah the chemins are similar that's a it's a different instrument but it looks at solid samples as well that's a seti institute nasa ames instrument actually the chemins have an x-ray diffraction system uh yeah and they've been very helpful as well and helping us label the data uh-huh So this is one thing, Megan, that I think you probably can expound on a lot is the whole issue of data preparation.
I mean, what is, you know, getting data ML ready or AI ready? It's not necessarily that you can just take a pile of data and throw a machine learning algorithm at it, right? There's a whole preparation process. Yeah, there is. And that was one of the big focuses that we had, you know, Eric mentioned the the dedication of the SAM team to prep the data. And that was a huge effort, you know, and labeling or getting it ml radiant.
And another way is I think, you know, being able to give a team at the beginning of the summer, something a little more usable was something that we really focused on this year. Yeah, yeah. Excellent. So just as a reminder, everybody, if you have questions and you're on the Zoom platform, I guess some of you are watching this on different platforms, but you can put your questions in the chat on Zoom and we'll be looking at those questions and having an opportunity to ask those to our panel here.
So don't hesitate to do that. Also, as we do with SETI Live and SETI Talks, we'd love to know where you're watching from. And so if you don't mind putting that in the chat, we'll share that information because it's always amazing how far reaching these sessions go. And we love to be connecting with people all over the world. So let us know where you're watching from and let us know if you have any questions. And with that, maybe we can talk about what has happened this year.
And you don't necessarily, if you don't want, because you don't necessarily want to give things away for tomorrow, say what results you got. But perhaps talk about the extent to which, well, what does success look like? And to what extent were these efforts successful, particularly in terms of applying the AI and machine learning techniques? Do you want me to start with that? Well, kind of our goal was just to sort of, first of all, is determine if we could do this transfer learning.
Because as Victoria was saying, we actually have this ironic problem that when we're on Mars, we have too much data. When we're on Europa, we'll have too much data. But right now, we don't have enough to do the training because our instruments are so particular. So if we could just prove that this commercial instrument develops data that we can use to train our actual space flight instruments, that would be a huge step to prove we could do that. or what we have to do to make that work.
And it looks like they've made, I haven't seen all the results yet. Victoria, you want to add to that? Yeah, I'll add to that. I remember the first day we met the team, Eric and I were like, it's okay if the result is, it doesn't work. At least we learned something. Yeah, no result is meaningful. Yeah, but It turned out that it seems to work, so that's great.
But also what we learned from this challenge is how to get the data preprocessed, like the data pipeline preprocessing that is always a struggle to understand what we have to do to make the data ML ready. could be applicable to other type of mass spectrometry data. So that's the fact that it's an eight-week sprint, like the data, every single day they work on it and they work as a team.
They ask the scientists, they were in touch with them, like enable it to be like used for other type of mass spectrometry for potential other challenges, other projects. So that's also a big effort that the team did this year. Yeah. So I think it's interesting because what we'd like to see in all the FDL challenges is the extent to which there's some generalizable outcome, right? That we can say, okay, we can take that approach and apply it over here or that methodology and apply it over here.
So you seem to feel that that was achieved. Yeah, to see that and also the fact the mentor, amazing mentor Robert and the other scientists and the team themselves They are able to understand this generalization process to help us also build future instruments, like what are their recommendations? What should we be thinking about? They have this long-term view also on the FDL eight-week program that is really amazing.
And would you say, based on what happened in this sprint, that you've made progress towards the goal that Megan's talking about, about integrating machine learning thinking into the mission planning process from the very beginning? some element of this that you think contributes to that mindset or that approach.
I think from that challenge for sure, but also from the preparation that Eric mentioned with all the scientists from the SAM instrument, from the KEMEN instrument, they are very curious about machine learning. It's not their field of expertise, but they are very interested and they want to know more and they ask questions. They want like, how can I make it better for next time? So they are also working with the scientists early on, makes it happen. Awesome. Janice, tell us about the moon.
Yeah, well, I would say that we've been applying, me working with our mentor, Mario Parente, machine learning algorithms for remote sensing on Mars for the last few years, sort of on a small scale. We haven't done an FDL sprint with that. But now with the moon, we felt we really need to catch up because there's so much more coming on the moon soon. And this M cubed data set has been kind of sitting unused in a lot of sense.
It's been analyzed by scientists, but it hasn't really had the benefit of machine learning approach. this summer we've been able to tackle that. One of the challenges is the low resolution global data set versus high resolution. There are a few or a smaller number of targeted high resolution sets.
And so we've been trying to work on running the data the code on the global data at low resolution and also building algorithms to train for the high resolution so that when you can map something at a global scale you can see differences between maria and highlands on the moon and differences between certain minerals but when we can go in and look at 100 meters per pixel resolution. And a pixel is a little, the small spot where you have each data point in an image.
So if you have 100 meters per pixel, it's way better than kilometers. And then you can start to see geologic features. You can see crater rays and you can see boulders and features on the surface. Not a boulder if it's 100 meters, but you can see boulder fields versus smoother terrain. And do you see a relationship between terrain and sort of surface morphology and the presence or absence of water or evidence of it? We do see some, yes.
And so we're looking at the geomorphology and we're also looking at the elemental chemistry and the mineralogy. But yes, I think the grain size of the terrain and if it's mature or if it's altered, if it's fresh, then we see it. We see differences. It's nothing worse than immature terrain. And so I know this wasn't what you were trying to get into this summer, but what's the deal? Enough water on the moon to support as a resource human long-term outposts?
I don't think we'll be building swimming pools or... But it's much smaller than on Mars, but there is water there. And I think what's really exciting is that we have both this... really firmly connected bound water and we have this mobile component. So I think part of what's coming out of this is scientists are going to need to rethink. Some people have talked about the possibility of an exosphere so that there's no atmosphere on the moon.
So where does that water go if it leaves the surface and then comes back? So some scientists have proposed there's some kind of a near surface region that these water molecules could be inhabiting and sort of cycling. Yeah, it's that they're connected to the surface and not connected. Because I read one of the write-ups you had that mentioned this recapture notion of volatile water.
And that's something where the science community is going to have to catch up with what the machine learning community is telling us, in this case, with the data. And so there are these two forms of water. So the bound water is going to be less easy to harvest, but this exchangeable water is going to be easier to capture. So it's still probably not going to be handfuls or teaspoons. It'll be little bits of water. But I think especially this, we're not looking at the poles and the icy regions.
We're looking globally, mostly around the mid-planet, not at the poles. So once we do get missions to the poles, we'll be able to see more. And Trailblazer will be coming on soon and using a similar technology. I think it's really exciting that all the machine learning that our team and Mario have been working on this summer is going to be applicable to Trailblazer. So that'll be really exciting to have that tool.
Well, you mentioned this idea of the science catching up with what the machine learning data is telling us, which is a great segue into the question I want to ask Megan now, which is about the acceptance of machine learning and AI as a technique. you know, the scientific community, and that includes people at NASA, you know, over the years, maybe it's become less since we started FDL, right?
But there's been, you know, some level of skepticism, like, to a certain extent, this is a black box, we don't necessarily understand what's happening in the black box, therefore, can we trust what comes out of it? Where does that stand now, you know, at NASA, and maybe more broadly, from your perspective in the sciences? So I think it's still a challenge that people face.
I think what's been interesting over the last five or more years is the transition of machine learning models to be more supportive of an enabling of scientists. So it's less about can this compare to a really robust scientific result using more traditional techniques and more asking the right questions of how to use AI and ML as a tool. for science, where, you know, maybe we don't need a very precise result with an uncertainty on it, right?
Maybe it's more of a, you know, a large language model helping you scan the literature and find knowledge gaps or something like that, right? And that, I think, can affect a scientist's life every single day. It can make their life easier. Even like summarizing notes from a meeting, you know, it can be as small as that. And we have a large language model that we trained through the Science Mission Directorate. So it has a little bit more understanding of the science that NASA does.
So if you ask it to summarize meeting notes or something like that, it can provide a more accurate answer, right? So I think there's been a transition to have it be more supportive of scientists. And I think that's where you gain the trust of these models because it is actually affecting their life every day and it's helping them. That's where I've seen the biggest change. Yeah. So there's a recognition that it's helping them. Yeah. Yeah. And maybe in the future, it'll be more science focused.
And I think we are heading in that direction. That's my hope. But I think in terms of building the trust in the community, I've seen that make a really big difference. I would say it's going to become more and more important because with missions to Mars or the Moon, they're close by. We get that data pretty quickly back and we don't have the huge data volume challenges. But as we go to Titan and Europa and other places, they can't send everything back.
We're going to need to, like you said, build it in at the beginning. The ML needs to be part of the foundation so that So we're making the right choices to send the right data back or go to the right next, if we can't, if there's no time for the scientist to be telling the rover or the lander what to do next. And it's kind of interesting, the decision. So I'm the program scientist for this thing called the Habitable Worlds Observatory, which is the next astrophysics flagship concept.
It's a couple of decades in the future and everything. But one of the things I'm really interested in is keeping this mission AI ready. So we're in very early mission concept development phase. Victoria has been very involved in this too. And, you know, what decisions are we making now that might compromise things down the road? So, you know, something as like, will the spacecraft have enough processing power to be able to run these machine learning algorithms, right?
These are things you really have to start thinking now, right? How does running an algorithm on a spacecraft are going to affect like the thermal issues, right? It's amazing the things you want to make sure don't affect things down the road so that you can do these things, you know, because flagships take decades to create, right? Yeah. You've got time. That's nice.
So before I ask some more questions, let me put a question to Rebecca and Beth and let us know if you've got some questions from the audience that we could address up here. We do have a few questions. Okay. I think Janice kind of touched on this and I think it would be a good one to sort of touch on again, but is there any water on Mars? Oh, on Mars. I would say there's a lot of water on Mars. There's a lot of H2O on Mars, and it's in different forms.
And it's going to be frozen on the surface at the poles. And there probably is a lot of H2O in minerals, in clays and hydrated sulfates. And there's probably a lot of H2O ice embedded in the regolith. I don't know if Sam's able to determine anything about water. I mean, it's sort of a simple molecule. You're mostly interested in large complex molecules. We're looking for larger complex, but there is plenty of water on Mars. It's not that easy to get it out of the state that it's in.
uh the so it sounds great that there's this water on mars but you can't just like drill a well and and get it out it's a it's an expensive uh it takes energy to get that water out so yeah so there's there's more and more finding wasn't there a recent finding um radar-based finding of or potentially of of subterranean water large amounts of potentially large amounts of subterranean water on mars yeah how to get it out is different matter another question yeah sure sure and now the other thing I i
heard about the the amount of water locked up um on the polar ice caps is that if it all melted now this the planet would be more or less covered with about 20 meters of water all around so I mean it's like a lot of water but to your point doesn't mean it's readily accessible or usable when on mars it's going to just um disperse disperse great interesting question um what else you have beth Sort of a similar question. Can humans drink water from the moon or Mars for that matter?
Is that possible? Well, I would guess it would have to be purified. I put it through a Brita filter. Like the Martian movie. He was trying to recycle water for his potato farming. But I think you know, like going, you know, to the mountains and drinking, you know, stream water, it's probably clean water, but you never know. It's probably good to filter it. And on Mars, it's not just, as far as we know, there's no life. There are no humans that might be peeing in your stream water.
but there there certainly could be you know acidic water or salty water really saline so there could be a lot of things that might not be super healthy for you so water is water but you don't know what might be yeah the h2o would be the same h2o on earth and h2o on mars or h2o on the moon that's all the same it's all the same Yeah, I mean, when you go hiking, right? We all know this. You have to be a little careful because you don't know who pooped upstream.
But that won't happen on the moon or Mars, to the best of our knowledge at this point. But, yes, other constituents could be there. That's fun. What else have you got, Beth? This is a good one. Again, same thing. Everybody's very interested in the water. Does moon water have a different deuterium ratio to that of water on Earth? Right. That's a good question, and I'm afraid I can't answer that right now. I haven't checked that. Yeah, I haven't thought about that.
I don't think we have the means to check that easily. Well, and the deuterium would, the spectral band would be shifted. So the water band that we're looking for at the moon is H2O and OH. And the so a lot of people studying trying to like from my work, trying to look at the spectral features for specific minerals, if you're trying to confirm a band assignment, say yes, this band I'm looking at is indeed a wage, then you change it with deuterium, because then it shifts to a different wavelength.
And then you can say, Aha, I was right, or I was wrong, and then figure it out. But I don't know the answer to that. People studying meteorites often look at the deuterium level, and I don't know. For Mars, it's a little different, but for the Moon, I don't know what the deuterium levels are. But they're not responsible for this water band that we're seeing in the M-cubed spectra. Got it. Okay. Another one, Beth? At this point, it's getting a little silly.
Well, I've got one that would relate somewhat. So we did prove, I think, scientifically that some years ago that the moon is not made of cheese. But you're doing mass spec data on Mars. What's Mars made of? Good question. It's made of – I mean, a lot of – I'm not the expert on what it's made of because we're looking for – We're looking for organics and complicated, complex organics. What we're finding, I mean, we're finding the minerals we expect to find, the rocks we expect to find.
It does have a lot of rocks that are similar to what we would find on Earth. I can't remember the names of the minerals right now. The SAM labels mainly are silicate, phyllosilicate, sulfite, organic molecules.
that's we have like 10 labels that will be presented tomorrow but that's really more the mineralogy and the evolution of this mineralogy throughout the where we drove on mars with curiosity and we have this upcoming mission with exomars that is a collaboration between issa and nasa with this instrument that Eric has been working for a few years now.
That is a MoMA, the Mars Organic Molecule Analyzer, that is also going to give us so much more information on the Mars surface and the Mars subsurface because it will have a drill on it. So we'll be able to drill two meters deep down. So that's revolution. And the type of mass spectrometer we have on board will be able to really study organic presence. But that's Long story short, we're looking for organics when we study Mars.
I think one of the questions for Sam has really been the nitrates versus the perchlorates or other chlorine-bearing things. And that's been so tricky. And has your team been looking at that at all this summer? A tiny bit in the way that we gave them two types of mass spectrometry data that Sam collected, EGAMS, so Evolved Gas Analysis, and GCMS, Gas Chromatography Analysis. So basically the big picture is from EGAMS and the more precise picture is with GCMS.
So that problem that you mentioned is more present in the GCMS data that we processed the team processed in a machine learning ready format, but did not go through all the models. So our focus for this challenge was more on the big picture EGMS data. So maybe a future year challenge, FDL challenge for GCMS data. And what's the latest with methane on Mars? Oh, that's I think Mars 10 conference. The latest with methane. We still don't quite understand.
The storyline was we got there, and we have it on SAM. We have a tunable laser spectrometer. It's very good. It was specifically made to look for the methane. Because before 2012, there was all kinds of wild ideas about the methane on Mars. And we were there for months. And you probably know this story. And found almost no methane. It was just at the very bottom of its ability to measure methane. So the story was, no, this enormous amount of methane we thought was on Mars wasn't there.
But then suddenly, a methane appeared. And we thought, we must have done something wrong. One day, there's 10 times more methane than the week before. We must have messed up. So we did it again. And it was still there. And we did it again. It was still there. So we had to say, yeah, all of a sudden, there's methane. And then we wait a month and we did it again and it was gone.
And this happened, occasionally it would come back and then it would go away and it would come back and it would go away and it might be seasonal. We've had enough after 12 years, I think it was 12 years on Monday since Curiosity landed. There's been several cycles of six years, I guess, six Martian years. But there isn't a really close correlation to the cycles a little bit, if you look at it right. Although I heard recently a presentation that It's very local.
We think it's very local because global observations of Mars don't see this come and go at all like we see on the surface. So we think it's right there where it's coming and going right there where the rover is. There was a suggestion that it's underground, and when the rover drills and moves, it actually cracks this kind of thick layer on top, and that is actually releasing the methane while we're driving.
I mean, it's just a a theory I'd heard recently which is very interesting that would be interesting in terms of where that methane is coming from well I think methane is often trapped in clathrates or could be bound to some mineral I think it's produced when olivine reacts and with some other minerals so I think it's not uncommon to have methane in this so it's just strange that it's coming and going but I think I i would guess rather than like little microbes that are in there
you know spitting it out from time to time I think it's more likely that the rover activity somehow unearthed this Yeah, the rover itself caused the release of the methane. It was trapped there in the regolith and that the rover activity released that from time to time. But why? How come one measurement one day released it and another day it didn't? And that's, I think, a good question. That's an interesting challenge when we start talking about things like edge computing, right?
I mean, decision making about what data to send or not send. And if you've got a long trend of nothing happening, then you stop looking for that thing. And You know, you're going to these places where phenomena may come and go that you don't expect, or you don't expect this time series dynamic to be at play. And then, you know, you might be making mistakes and not sending interesting data at a given point in time. I don't know, but it's really challenging.
How do you see, in terms of this, vision and notion of integrating machine learning into mission planning. Where does that stand now at NASA? Is this the very beginning of this endeavor or is there a momentum in this direction? What's the latest? Well, there was an article, a press release that just came out that talked about some machine learning algorithms that were put on some of the rovers on Mars and using that to do onboard autonomous detection and things like that.
That was the first time that I had heard about it actually being like deployed. Okay. But yeah, I think, you know, getting back to the idea of trust in machine learning algorithms, you know, these missions can cost hundreds of millions, if not billions of dollars. And you want to make sure that these things, you know, that people do have trust in this because that is a lot of money. But I was really excited to hear about some of these things being deployed on some of the rovers.
That to me was very exciting and a very big step towards it becoming more common on other missions. Yeah. Well, I think just to add to that, if we start the mission with the machine learning in mind, the scientists will go through the whole mission hand-in-hand with the machine learning, and they'll learn to trust it because it takes six to eight years to build a new instrument.
And they'll spend those six to eight years, and hopefully that whole time they're seeing the results of the machine learning. So they'll kind of grow comfortable with how it works. And to add to that, also, there is the trust side, but there is also the hardware side, because a lot of stuff we do on Earth is because we have the hardware on Earth that the computing power, all these sometimes we don't have on the rovers on the mission.
But I know NASA is investigating other hardware processing to enable this machine learning to be on board also one day. And I would say with the Rovers as well, that you usually have several instruments and these instrument teams work as a team to address their data and resolve their data. And they're usually machine learning or whatever to get to what their data means. But then you've got seven teams or something or 10 teams or whatever. and like how Sam and Kevin are working together.
So if you can build it into the architecture that there could be cross team machine learning, that would be so helpful because the current in the last couple of decades seems to be that each team works first until they understand their data and then they broaden out to try to coordinate. And if they're coordinating from the beginning, they might be able to calibrate their data a little faster and get to some results that are gonna help the rover and the mission faster.
To add to that, when we prepare the challenge with Eric to get the data ready, we had workshops with the Kamin scientists and the SAM scientists. And you had these workshops ahead of time, right? Yes. That's amazing. We started maybe September 2023, so almost a year ago. And seeing how the team interacted, because they are used to work together. They are colleagues for more than 12 years now. And it was so interesting to see how they were actually working together.
Like, I see this in Kamin, so that means that in SAM. And implementing this in machine learning algorithm something I'm sure NASA is interested into and will be fascinating. Yeah. Yeah. I mean, I know one of the things our our science director is very interested in is data integration, you know, through through columns like what do we see from space? What do we see in the air? What do we see on the ground?
And can we integrate those data sets and learn more about systems and processes through those kinds of integration methods, which I think almost invariably would require machine learning to figure out and connect those dots. So getting back to the FDL program itself, so now you're all like veterans at one level or another. And so one of the interesting things about this format, and Megan talked about it, is it is a sprint. It's an eight week, very intense, heads down, hands on process.
But it does seem remarkable over this timeframe that the teams get results extraordinarily. And so I think, again, back to NASA's questions, surely seems like you can have interdisciplinary teams of computer scientists and domain scientists working together and doing extraordinary things in short periods of time. Has that been what you've observed from this? And what were you expecting at the beginning?
I would like to jump in and say, I think the power is having the scientists working with the machine learning part of the team. So it's not just, okay, you go off and do your thing or you go off and do your thing. And working the last couple of decades with engineers, it's almost like a foreign language. And it shouldn't be that way. You know, you young people, you know, you're probably changing the world so that it's more interrelated.
But I think FDL really sets up a program where we have scientists and machine learning experts working together and they're talking every day and they're addressing each aspect of the program. And I think that's what's needed. And that's the real magic that FDL offers that you don't see in other programs, because usually if you need something done, you just hire it out. It doesn't work for this kind of a problem.
Sure. Yeah, I like what you're saying because I usually say that my work is being a translator between the science, the machine learning, the engineering, the management, being able to talk to each subset of people to make sure we are moving in the right direction.
So in that team, the Mars team, I'm not sure it's the same for the lunar team, we could really see at the beginning, Eric was there the first week, I was there a few weeks after, how they were... trying to explain to each other what they understood from the machine learning side, from the science side. And just before this meeting here, we were practicing with the team, and they all seem expert in machine learning and science now.
So it's very impressive to see how much they learned and how much they understood the problem, the initial problem, and machine learning as a tool, which what did they try, what worked, what didn't work. But they can all talk about the science and all talk about the machine learning after eight weeks. Which is great. You know, it's interesting because you can well imagine that we are turning scientists into machine learning experts.
Are we turning machine learning experts into scientists as well? And I think there is some of that. One of the early years, I can't remember if this is your year 2018 or not, but there was a woman named Ann Maria, who was a computer scientist who was doing like economic modeling and forecasting using machine learning and AI for like banks and things of that sort. She came to participate as a researcher on the compute side in an FDL program. She's never gone back to banking and finance.
It's all about space. She shows up everywhere now. Yeah, the science is much more fun. So any closing thoughts?
One of the things that we'll actually talk about amongst us more over the next day or so is what's next and how can we let's say modify or enhance the fdl model to provide more continuity in in the research that gets done so there's a you know what what comes next because I know that one of the things that's been frustrating for us for the teams and for nasa has been you know once the sprint's over and it's over and yet sometimes it's like there's a really promising result here we'd
like to you know go to the next level so um I want to talk to all of you a little bit about that how we integrate that kind of thinking into the program and processing but talk about um you know what you see as kind of an evolutionary next evolutionary step for a program like this well I'd like to well you know victoria this challenge came out of previous challenges that you did not fdl but this was an evolution and so I think just having that view of this is the project for
this year but you know it can lead into other things um I i don't know I think that really helps set up this whole summer with the things you you both did beforehand yeah yeah And I remember when we first met with Bella, Bill, you, Eric, and I to potentially have that as an FDL. We were like, we had so many ideas. Like we say, oh, that could be FDL1, FDL2, FDL3, and have a continuity over the years. Because this is just a big question, like machine learning for mass spectrometry data.
It's so, so much question. We have, I think, way more questions now at the end of the challenge that we had at the beginning because that seemed to work. So how can we improve it and answer it? How can we implement it? So there is that part of continuity for sure that is very interesting. Science is incremental in a lot of ways. Why not science and machine learning, right? 100%. Why not view it as the same? Yeah. And actually, more questions is better, right?
I think that we know when the science is doing a good job when each question we answer raises 10 new ones, right? That's the fun part. Yeah, and also ideas. I think the biggest thing that surprised me is there they came up with, the team came up with more ideas than they could try. Yes. So they have a whole table of, oh, this is something we would have liked to try, but it was going to be a little bit more effort. We needed another week or two to do this one.
And so we have a list of these, which is exciting. Which is good. So that almost is a built-in what's next. We can talk about those ideas. We had that for the lunar team too, exactly, that they're raising more questions and more interesting science directions that we're going to work on. Yeah. Victoria? Yeah, too, because you say concluding all this.
So I think Eric and I, we really want to thank NASA for funding these projects, but also the SETI Institute for Bella, Mario, all the... all the mentor, Robert, to make it happen because I think they had a lot of fun. They learned so much. And again, all the SAM scientists and the NASA Goddard and Johnson Space Center scientists that helped us get the data ready because that was a huge amount of work. And the team could work on the data, like starting FDL.
So that was really, really important for us. that's wonderful all right thank you very much well it's probably time to wrap it up I don't know if we have some some location information that we want to share here I don't even know where it would show up at this point So just to welcome people in, we had people watching from all over the world, from Australia, from here in California, from Sweden, from Hawaii, Missouri. There was someone from Ethiopia, I believe. And so just all over the place.
It was a very well-traveled global audience. That's excellent. I've heard of Missouri. Yeah. That's excellent. And we do want to say, if you're listening, we should say hello to Victoria's mother because hopefully she got word in time. Thank you. So with that, we're going to wrap it up here from the SETI Institute. We appreciate your watching and being with us today.
Just as a reminder, the SETI Institute is a nonprofit scientific research, education and outreach organization located here in Mountain View, California. And if you'd like to learn more about our work, you can visit our website, which is www.seti.org. And you can sign up for our newsletter. There's kind of a cool science weekly newsletter, our e-news that comes out every week. So sign up for that. Learn more about what we do.
And if you are interested in learning more about the FDL program, As an interested member of the public or as a potential participant, go to FDL4AI.com and learn more about that program. And again, after tomorrow's readouts where we'll get the real news, which we couldn't reveal today, we'll be posting those videos on that website so you can come back and watch if you're interested. It should be a lot of fun, and we're looking forward to tomorrow.
And thank all of you very much for this conversation. It's been a lot of fun. Thank you. Thanks, everybody.