 But it now gives me great pleasure to move on in the program to the second of our two keynote talks. And we're inviting Dr. Scott Osking from BAS to give our present our second talk. So Scott started off in environmental AI back in 2016 and with a few papers bringing novel digital methods into the research that's going on here at BAS. And in 2018, Scott set up the AI lab, the renowned AI lab here, supported with a bit of funding from the Alan Turing Institute, the ATI. And then, in fact, Scott, you joined the ATI in 2020 with a, that's a comment to sort of expand that group and build the links. In February 22, Scott became the BAS science leader in AI and joined the BAS science executive. And now I think Scott has 20 colleagues reporting across both institutes, including scientists, engineers, and community leaders. And obviously we've met some of those colleagues this last presentation this week. So this month, Scott became the co-director for the 6 million pound Turing Institute Research and Innovation cluster and Digital Twins. It's hosted by the Alan Turing Institute where Scott's going to expand the research activities and develop further spatial AI, spatial temporal AI methods for environmental digital twins, which is such an exciting area. Scott, delighted to have you. I didn't have to come very far. Thank you, Stephen. And thank you everyone for coming and holding out today too. Welcome to BAS. I think we've said that already. I want to mention that the building that you're in now was designed, or the architect was designed, inspired by a glacier. So on the way in, you'll see the hexagons and the shape, the angles. Also, if you look at the ceiling, this is inspired by the ice cores, the bubbles that come outside, trapped in the ice core. So I thought I'd give you a bit of trivia to set a scene and where we are today. So, Stephen mentioned my background. I'm a climate scientist by background. I've been moving in to bring it in machine learning AI methods into BAS and to the group. And one thing we've done there is to stand up a cohort of machine learning AI experts, rather than distribute them across the building, it makes sense to those people to be talking to each other day in day out. And I'm going to explain some of the work we're doing in that space. Just to help motivate things, this article came out in the MIT Technology Review just last week. Eric Smith, who is a big name in Google and tech, as hard as how AI will transform the way we do science and how science gets done. And the very first thing he kicks off with is this paragraph, which is focusing on the unprecedented heatwaves, wildfires that we're experiencing right now, the flooding, you know, the climate crisis, biodiversity crisis that's facing the times, and how AI and digital twinning technology could support either how we mitigate and adapt with our changing climate. In terms of what we're doing at BAS, that connects to that idea, that vision, we're a highly multidisciplinary institute. We focus on the ocean, circulation by geochemistry, glacial change, the geology of the underlying continental shelf and the ground in which we work with satellites, data, we fly drones, you've seen our ship a number of times, we have bases and manage those as well. So, highly interdisciplinary, we have a space weather, we had an excellent talk yesterday and so the project, you know, we cover anything, anything up from space weather right down to the creepy crawlies on the ocean form of benefit biology. So highly multidisciplinary, this makes a really fertile, you know, group, if you like, to be running AI art, we have diverse datasets. A lot of those datasets should be interrelated, although they're all measuring the same system, so there should be a lot of correlation or relationships between variables. So the big question here is how do we combine these datasets together and I like to see BAS as a really good playground, but you know, we put all the NERC centres and all the NERC labs and environmental labs and university groups that you represent provide a really good playground for AI and machine learning. The focus of this talk, I added the word spatial temporal modelling because I could talk about all sorts of things that I thought let's provide some scope here. Spatial temporal modelling is important throughout, you know, all the sciences that we've touched on over the last couple of days. Spatial temporal modelling, how, you know, stuff moves in space and time, very difficult problem, you need to work at scale and, you know, if you're interested in extreme events, you really do need your models to be accurate and have robust uncertainty. So I think that we're focusing on in this space includes the fusing of different sensors from satellite sensors and surface sensors shown here on the left. We're fusing, we're working with fragmented datasets, you know, due to clouds in satellite products or maybe one of our sensors as covered into those we have gaps in our data, we need to be able to find ways to integrate fragmented data with more sleep data. And then often we need to increase the resolution or get to those granular scales, those impact scales, you know, using super resolution or downscaling methodologies. Here's one example again where we need to get down to those local scales, but also we need to optimize our, the way we do our science, the way we take descent, we place our sensors or the way we run teacher climate modelling experiments or ice sheet modelling experiments. So here's an example of our ice sheet model, the WAVI project here which is part of our sea level rise research here at British Antarctic Survey. And often we run these models forward in time, where let's say we're interested in sea level rise contribution, we might start with the current state, and those models, those climate models or ice sheet models, they evolve in time. And ultimately what we need to know is what the uncertainty, what the predictions will look like at the other end on the time horizon that you're most interested in, let's say a hundred years or a few hundred years. Now we cannot keep running these high resolution models, these cloud models indefinitely, they need to scale up, we need to find optimal ways to identify where we should run the next experiment, you know, what resolution, what physics should we be set in, what physical constraints we're putting in the model to optimise that, that uncertainty to get the best robust prediction window at the end. So this is again where AI is coming in whilst social terms modelling is a hugely difficult challenge but something we need to address head on. Another nice example here from a few people in our group in Magic, the mapping group here working on is sea ice and ice cloud products. So here we have three different satellite sensors. We've got the pathway of microwave imagery down the left hand side, which the resolution is of six to 25km resolution. We then get down to another satellite sensor, the visible imagery down to 250m resolution, a daily temporal and a daily temporal coverage, and then down to these 40 or 80m resolution, still trying to measure the same thing but a much finer resolution, 40 to 80m there. And one of the challenges now we want to combine these different data sets to get these different products because they all have their own strengths and weaknesses and we need to find ways to incorporate them all together. Yeah, just to follow that with one more slide. We're not only interested in the static imagery but we also want to track changes over time. So Ben and others in our group, Ben Evans is here today, looking at tracking of icebergs as a continent. The icebergs provide a really important metric of the loss of ice from the continent and to help us project that seal and rise. We're interested in the make up of the sea ice and satellites and you can see here, you know, when you get down to these really fine resolutions, you can see these ice flows, these blobs of ice, and we can use machine learning and deep learning computer vision tools to segment those ice flows. And again that's hugely important for a number of reasons which I'll just quickly go into before we launch into the main part of the tour. It's also important, you know, you might think it's miles away, I've never seen it, why do I care? It supports the native people and the wildlife and the communities that live across the Arctic and the wildlife as well and Antarctica as well as the sea water itself in their fishing, hunting and migration activities. It also drives the ocean currents and circulation as the sea water freezes up, it rejects the salt insulation which drives the circulation and a lot of our climate as well and then the planet. So that's the mirror reflecting a lot of sunlight back to space so as the sea ice reduces the dark ocean absorbs that heat and warms up more quickly and also influences our weather even in the mid-last shoots as well. But also the sea ice is hugely important for navigating our ship and autonomous vehicles which is why we also spend a lot of time thinking about how sea ice is going to change what that means for our polar operations over the next few months or next few seasons. So we've just launched our new strategy and, you know, just to show how important sea ice is, we've stuck it on the front cover. It does, sea ice really does touch pretty much all parts of the building and our biology, our polar operations, our climate, our oceanology. So it's a hugely central component of what we do. I mentioned also where I had at the Elm Turing Institute, so as well as launching this strategy, you know, it was involved in helping to write this strategy, I've also been leading the Turing's environment sustainability grand challenge and their strategy in the space. And the areas that we're focusing on there is the automate by diversity monitoring to enable nature recovery. So utilizing infrastructure to enable use of natural resources, modelling interventions to achieve sustainable cities and regions for a net zero world. And then the one closer, I guess, to my background research, delivering localised environment predictions to mitigate impacts of climate change. And I wanted to highlight, you know, I do have these two hats, Steve mentioned, I have a hat here at Elm Turing, but I'm trying to merge them as much as possible because there's only so many hours in the day and I don't want two jobs. So the idea is to really start finding ways to optimally work with, you know, these, these AI communities, the area of groups and the funding that's available there and how we can also bring that into the NERC, the NERC community and the NERC system as well. Yeah, so big highlights from the AI lab is this ice net model and seasonal CIS prediction system that we built a few years ago. Some of you may have heard it, I pretty much mentioned it in every talk, so apologies if you have already heard this. But the idea here is that we want to predict how CIS will change in the next few months over the next season. We input multiple weather variables, winds, temperature, pressure, et cetera, et cetera. Well, the novel aspect of this model is that we pre-trained this model on climate simulations. So climate simulations from the past and also to the future up to 2100. So the model has seen something of the different climates that we might experience, not just the satellite data over the last few decades. And then we find tune with observations as well once we've done that pre-training. The take home message of that project is that we could outperform a state-of-the-art physics-based model on lead times forecasting head between two and six months. And we did this within 18 months and the State of the Forecasting Centre won't name them here, but they've been working on this for decades. So it's a real game changer. AI, going back to Eric Smith's comment earlier, it really is a game changing time for the way we do environmental science. And also I should say that this, now this is all trained, it takes 10 seconds to run on a laptop. So, you know, think about real-time decision-making. This is, you know, clearly has an advantage to use. So once we've done that, we wrote a paper, we have loads of citations, it's been highly impactful. Let's think through this, the process by which we got there. So we came up with an idea, we did the research, and then we wrote this paper. Now, historically, you might imagine, okay, so you've done this, then I go back to square one and write another research proposal, come up with an idea. And we didn't want to do that. We really wanted to take this, you know, state-of-the-art series forecasting system out to market, if you like, to put it in the hands of the people that need it. So we are now working on a couple of projects, but this is one example, a project that I did about in our group's work. You know, the caribou or the reindeer here rely on CIS for their migration, for their, you know, the seasonal migration following the CIS and looking at the seasons. And what we've done is we have, we've got data where these caribou have been tracked, they've got collared data, GPS collars on them, and then we can track how they move in relation to the CIS change. So we have to use that CIS change to help understand and predict, you know, the risk that this community, this population, you know, might be under the next few months. So as you can see here, the CIS is starting to freeze up, as you see more yellows and green colours, it starts to freeze up. And when it reaches a certain level, you'll see these caribou start to make a dash for it across. This is, you know, any really cool animation, but B, it will provide us now with some indication of what CIS concentrations need to be reaching, how frozen does the water need to be before the caribou are constant enough to step on the ice and migrate it faster. And if we can predict whether we're going to get a low CIS season, we can feed that information back to the local conservation groups. And then also building on this is something that James Byrne is working on at Bass is the way to alert these communities. So we're working on mobile, you know, communication tools to send out alerts to those who need it, again, to get the forecast and information to the hands of those that need it. So working with these groups has been fantastic and really eye-opening because one thing that we've, you know, keeping in feedback on is that the resolution that we're working with is currently too close to make the kind of informative decisions that they need. We're now working on a super-resolution AI approach, a resolution imperative approach to get down to the resolution needed for that caribou use case, little system, others that I'll mention in a minute or so. But yeah, we're looking at a four times increase in resolution and eight times increase in resolution. So what I've just described there is we've monitored, we've observed the physical environment, the physical system, and we're using that to understand the system and passing information back, passing data back to a digital model. The digital model might be the ISNET model or it might be the iSheep model, the YV model that I've mentioned there is the CI level contribution. But what we're also doing in our group now is we're building digital models and we're passing that information back to the sensors or to the ships of autonomous vehicles. So I'll give a couple of examples of what we're doing in that space. Antarctica is very difficult to get to, highly hostile environment, we need to maximise, you know, when we go there, it costs millions to get there, when we go there we want to make sure we're putting the right sensors in the right location. So one question we have here is where should we place our new sensors to maximise their usefulness and our understanding. So we need to think about that when we put up our autonomous weather stations, but also where we fly our mobile sensors, you know, our drones and our underwater vehicles. So we've just published this paper or it soon will be published but you can see the preprint online. We've got this new project called Deep Sensor that Tom Anderson and my group is leading. My idea here is, you've got your contact points you've got your locations your sensors already, but you also provide some auxiliar information which might be the elevation of your mountains or the slope the angle how close these stations are to the coastal regions. The, the, the context in which these locations sit with large pressure systems, all that information is required to then make a prediction at another location. The temperature drops off as you increase in height so you know we need all the auxiliary information if we're going to fill in our gap data gaps and make measurements at other locations. So this is the, for the geeks amongst you who like to see a network design this is our machine learning approach. Let's show how we're using that in the real world so we've got this is just a sort of toy experiment but on the bottom left here we have some sensors already known to have to get these are temperature sensors. And as we go from left to right we are adding on each time a new sensor the sensor in red is the one that we include we're adding on and the shading here represents the reduction in uncertainty that that sensor has provided us with. And you'll notice that we added a red dot here and you might imagine that around that location you do see a reduction in uncertainty. But there are other locations that around here where we've also reduced uncertainty as a result of having that sensor, because the the climatology the weather and the environment up here has some, you know, little similarity to what's happening down here. So now suddenly when there's like environment where if you're interested in one environment, you don't necessarily have to stick a sense that maybe cheaper to stick a sense of someone else in a very similar environment and then use AI to reduce your uncertainty across the continent. So as we go from left to right, we're adding more and more sensors. And, you know, this is the optimal way to monitor the whole continent, but we're just for sensors. So this is our toy experiment here. But we're also now talking to CEH and Matt Fry thinking how we also deploy this make sure it's generalizable for other types of sensors. So here for soil moisture, we're also working with our quality sensors as well. So there's a really, you know, I see the Polo examples are first case study, case study zero, but every time we build something it needs to be generalizable and work beyond that one environment. Also for automation. So Maria Fox, who's here today and Johnny and others, and George here, and I've got to mention everyone and Michael's here as well. And James is involved. Hopefully I've not missed anyone, but we've got a team here, quite a substantial team now building and using AI for automation. The question here is what is the most fuel efficient route and task sequence to complete our science objectives again we've got this net zero goal to by 2040 2050 and running a ship is hugely carbon intensive so how can we make the best of the assets we've got. We're not just talking about the ship we also need to optimize the sensing and the autonomy of our drones and water vehicles. So this is just a nice example I've taken from one of Johnny's slides that shows a battery optimized routes between a and B between rather than Stanley. The, this is a battery optimized glider so the the route that we will take, let's say going north, the blue line here will be different to that coming south and that's because we're benefiting from the ocean currents to wins the that we're in so we're not just going from a to B we are gliding with and, you know, taking advantage of the of the environment as well and again that just shows how we can incorporate our CS prediction systems are so we're there also as part of this project need to forecast ahead in weeks and the CS is going to change and again you can take that all into account for me when we make these routes. So now going back to this idea of information between the two we now have a physical system out of digital system or digital model, and we want to pass information in both directions and at this point, you might argue that we are starting to develop something that you might call a digital pathway where under the correct time time horizons that you're most interested in your your physical system is pass is taking information maybe in real time or maybe you know that's passed on through a slower method but that's updated with your digital model so then that's a bidirectional flow you might call a digital twin and we're not. Of course, I'm dealing with doing that as a massive project was Europe for destination earth with ESA and WF and others are building a digital twin with ambition is to build a digital twin over the next 10 years. For the whole planet, that's the monitor understand similar similar and anticipate environmental change. Now with my chewing hat on, we are also spinning up a big digital twin efforts here in the UK. So this is the digital that during research and innovation customers digital twins and bass. It has a part in that now but also a lot of other net sensors were starting to talk to others now to grow that activity. So, and touch digital twin in well include you know the times, you know, innovation data sensors, the ship, but also where where this is somewhat different to what basketball do on its own, is that we're now looking at the how this underpinning technology can be generalized and worth across different systems. So the Turing already has a digital twin of an underground farm and building a digital twin of the heart and aircraft and the idea is that we don't we need to stop duplicating you know when we build digital tech. If we can build something like such as a spatial temporal model, we shouldn't just do it in the community, we should also find ways to work more closely and make, you know, a great high impact without without colleagues that are trying to do essentially the same thing just in another domain. The, the trick the digital twin center was launched only in March and actually the key highlights that the Turing Chiefs, Chiefs scientists demonstrated was the combination of our CRs forecasting system with our share. So, it is one of the current good examples of how we're progressing towards the digital twin, and essentially it's the integration, you know, different digital systems together. And certainly the integration of two systems, a CI system and ship when they're starting to think of, you know, expanding that and hopefully James will forgive me sharing this slide of his little old man, but we are integrating ice net there into, you know, our ship digital infrastructure. The A, the CETA Ratchable ship itself and sensors we have the digital version of that ship, maybe when we call it digital twin and not deductive but it's the digital representation and then on the left hand side we have the digital twin of, you know, the environment of the I mentioned earlier. So you can see that we're starting to bring all those components together, and I'm just running out of time, so I'm going to finish now. So in the digital space, we're in a really interesting place to integrate and work across disciplines across the biodiversity climate domain infrastructure, and to leave you with this final thoughts from the conversation. So Nobel Prizes most often go to researchers who defy specialization, winners are creative thinkers and synthesize innovation from varied fields and hobbies. And I cannot think of a better community in the environmental space to do that than the net digital community. So I'm going to leave you with that, and I hope in 10 years I'm celebrating someone's Nobel Prize in this audience who had it here first. Thank you very much. Thank you very much. Not fascinating. Please questions into Slido will start off with a question from Michael. Scott, could you comment please on how well understood the positive feedbacks and negative feedbacks are of some of the natural systems that you're monitoring sort of feedbacks positive and negative. But I'm not sure entirely understand the feedbacks between what is. So the, the ice sheet reducing reflecting less feedback that if the ocean wanted on more than more ice sheet methods versus like increase cloud perversion in more of the sort of radiation. Okay, so, I mean this these are big open science questions that we've been talking for a long time in clubs for instance the weather increase in cloud will have a positive or negative impact on Earth's temperature you know it is still up to debate the uncertainties are still large. The, the advantage of taking the digital approaches that we can start to plug those uncertainties look better so if we need more information over space or time or if we need to increase the resolution temporal resolution of our data and certain locations to understand the physics, let's say of the cloud or of the sea ice which are so important for for those feedbacks then that's where I think digital and automation can have a really important role. I get to ask you a question as well. I'm just interested from a sort of general points for learning, a little learning point for everyone here really is just AI has expanded the scope and the ambition and the scale of the sort of science that you're doing I wonder if you could try to characterize what, what the what the advantages of AI are to the general science. Yeah, that's an interesting point so I always say to our scientists here AI is not here to replace what we do but to get to do this stuff that takes too much of your time like you know that you'll see pie charts where 80% of a data analyst time is taken a fair approach to your data and automation can get hopefully a loss that will just disappear or will be reduced over time. So I think AI could allow scientists to spend more time doing science and that's, you know, I was trained to be a scientist to do science not to clean up data sets. Also, one thing we learned in the encoded in the lockdown was the speed at which we need to make decisions so you know we couldn't wait for months to understand when we should be locking things down or you know when we should shut down schools exactly we need instantaneous answers to things and as long as you quantify the uncertainty how confident you are uncertainties that's fine and I think, again a digital twin approach can provide the instantaneous response which may not be the most accurate response you might get if you went to a traditional model simulation but it might be good enough. And I think ministers, you know, more and more in the future going to ask or going to require very quick instantaneous answers when they have questions. Yeah, Scott, I think you possibly answered the sort of the next and last quick question which is just this conference has identified for four themes of next generation sensing data science tools and techniques the way we collecting data science and how do we bring all of this together into a package of contemporary modern science. What's your take on. Yeah, I think we still, I think in person meetings that is so important to keep those communications going and we did, we lost a bit of that. Over, you know, the lockdown period. So I'd like to see that this community keeps working together keeps finding ways to hack on data together, identify, like I mentioned earlier the digital twin of the heart and the the aircraft and the underground farm. So actually, when you sit down with them you realize you've all got the same problems might use different words for the problems where you have different timescales that you're interested in that we're interested in sea level rise over millennia. They're interested in crop growth over a season but actually when you distill it down to the data challenge actually there's a lot of overlaps I think, almost certainly within the with the interest I say we've been funding the same thing multiple times and one thing I'd like to do is to keep those conversations going and find out ways we can work on larger, more open source projects and you know, for us all to work together rather than working inside it. Scott, there's a lot of insight there. Thank you very much for your presentation.