How will AI change the way we understand climate risk? Listen to Jason Mitchell discuss with Professor Philip Stier, University of Oxford and Director of the Intelligent Earth Centre, about what AI could transform in climate modelling; where it can genuinely extend what physics and earth observations currently tell us; and, ultimately, how better climate information can lead to better decisions and better pricing of risk.

Recording date: 05 October 2026

Professor Philip Stier

Philip Stier is Professor of Atmospheric Physics and Director of the Intelligent Earth Centre at the University of Oxford. He also heads the Climate Processes Group and serves on the steering group of the Oxford Climate Research Network. His research addresses physical climate processes in the context of anthropogenic perturbations to the Earth system, the underlying cause of climate change and air pollution, with key interests in aerosol and cloud physics, their interactions and their role in the climate system.

Episode Transcript

Note: This transcription was generated using a combination of speech recognition software and human transcribers and may contain errors. As a part of this process, this transcript has also been edited for clarity.

Jason Mitchell:

I'm Jason Mitchell, CIO for Responsible Investment at Man Group. You're listening to A Sustainable Future, a podcast about what we're doing today to build a more sustainable world tomorrow.

Hi everyone. Welcome back to the podcast, and I hope everyone is staying well. So the big question for me is this, if AI can make climate models faster and more detailed and more accessible, does that also make them more reliable? Look, it's clear that AI has driven a Cambrian explosion across every field, but I think it's particularly interesting in the area of atmospheric physics and climate science more broadly, especially when it comes to how we understand and model climate change.

So if there's one thing you take out of this episode, I think it's the temporal dynamics at play. Here's what I mean. Traditional climate models are physics-based and equation heavy. They're vital for understanding out of distribution extremes. Think events decades from now whose magnitude we've never even experienced. AI, on the other hand, is powerful for understanding in distribution extremes like observed weather phenomena that's grounded in historical data and approaches like re-analysis.

Now, the linkage to financial markets I think is key here. Again, think about that temporal dynamic. Sustainable investors tend to approach climate as a long-term risk management problem, modelling the impact on portfolio value 20, 30, 50, even 70 years out. And to be clear, that's a necessary and important exercise, but it often pushes investors into unresolvable debates about discount rate assumptions or idiosyncratic views on cascading tipping points. The point here is that while that debate is ongoing, we're already struggling to price weather now. In fact, insurers are probably the best example. Just look at the premium inflation and coverage withdrawals we've seen following extreme weather events, which is to say that there's an opportunity to price the space between a weather forecast and a long-run climate projection to understand the extremes, not just the averages.

So it's great to welcome Professor Philip Stier, one of the scientists at the forefront of AI-driven climate modelling. We talk about what AI could transform in climate modelling, where it could genuinely extend what physics and earth observations currently tell us, and ultimately how better climate information can lead to better decisions and better pricing of risk. Philip is professor of atmospheric physics and director of the Intelligent Earth Centre at the University of Oxford. He also heads the Climate Processes Group and serves on the steering group of the Oxford Climate Research Network. His research addresses physical climate processes in the context of anthropogenic perturbations on the earth system, the underlying cause of climate change and air pollution with key interests in aerosol and cloud physics, their interactions and their role in the climate system.

Welcome to the podcast, Professor Philip Stier. It's great to have you here, and thank you for taking the time to talk today.

Philip Stier:

Well, thanks for having me. I'm looking forward to the chat.

Jason Mitchell:

Absolutely. So Philip, when I last saw you present at Oxford over the last summer, someone in the audience asked what climate models would look like in 10 years. And I distinctly remember you reflecting back in grad school, worried about a midlife crisis, that the opportunities for discovery and contribution would somehow narrow. But instead, these now feel like in extraordinarily exciting times to the extent that it's even hard to predict what climate models will look like in 10 years. So say more about this. What's changed? What research does AI make possible that you didn't expect to see 30 years ago?

Philip Stier:

Yeah, it's interesting you bring this up because to be honest, I don't know what happens in 10 years because things are changing so quickly. So just in weather prediction, climate models underwent a really continuous and quiet evolution over the last 30, 40 years, and they got better and better, but in a very continuous and steady pace, but they haven't changed dramatically for quite a long time. But then two things happened almost at the same time. So on the physical modelling side, it became possible to conduct really high resolution simulations that resolve processes. We previously had to parametrize, so estimate them from the resolved larger scales. That is one big development. And at the same time also, AI has rapidly transformed the entire field. So in weather modelling from start to the end, but also in analysing earth observations in very unprecedented ways. And now we have also hybrid physics and AI climate models. So where this all then goes for climate modelling in 10 years, it's totally up in the air.

Jason Mitchell:

Okay, so we're definitely going to dig into a lot of these things. So thanks for setting the scene. I guess first, your work specifically focuses on clouds and aerosols, basically small scale processes and more specifically tiny particles that can change how clouds form and how much sunlight they reflect. I guess for context, what's at stake here scientifically? Why are these interactions so important in climate modelling and also, frankly, so hard to pin down? In other words, why are these processes so critical to building more accurate models and reducing uncertainty in climate predictions, particularly around extreme weather events?

Philip Stier:

Yeah, so while everyone knows clouds, and many of us love them, many of us hate them, but very few actually know the importance for climate. Clouds have a massive roles in the climate system by not only delivering some of the impacts of climate and climate change like precipitation, but also by regulating earth's radiation balance, so the energy balance of earth.

And so clouds on the one hand reflect sunlight back to space. That's why they appear bright when you look at it from a satellite image and thereby they cool the earth by reflecting some of the sunlight. At the same time, they also absorb some of the outgoing infrared radiation and thereby warm the earth. But the net effect of clouds is actually to cool the earth by about order of 20 watts per square metre. So the number may not mean much to any listener, but order of 20 watts per square metre is about 10 times bigger the anthropogenic effect of CO2. So if we change clouds by just a little bit, we could actually entirely accelerate or double the effect of anthropogenic CO2 or we could totally offset.

So the question is then how do we change clouds? So first of all, clouds respond to global warming, so changing temperatures, changing structure of the atmosphere. But then secondly also because, and this again is another fact that very few people know is that each cloud droplet in the atmosphere is formed on a small particle. We call it an aerosol particle, and we change the number of aerosols in the atmosphere by air pollution because air pollution particles are aerosols. So by changing air pollution, we change the number of cloud droplets in the atmosphere dramatically, and that tends to actually cloud with the same number of water where same amount of water and more droplets tends to be more reflecting and reflects more sunlight back to space and cools the earth by about half a degree.

So that's the challenge why we need to get this right because on average, so when you think of half a degree cooling just by anthropogenic aerosols or air pollution, that is a significant part that could bring us closer to dangerous level of climate change, which have been defined as 1.5 degrees in the Paris negotiations and we are already there. Now if we were to remove our air pollution, we would be well beyond that and close to 2 degrees.

Jason Mitchell:

Yeah. Can you say a little bit more about this? Maybe drill in a little bit more because this reminds me of the conversation I had with Tapio Schneider at Caltech who you worked with last year. And specifically, I think people are aware of the reflective benefits of clouds, but I think more specifically when it comes to climate models, it's these small scale clouds, I don't know if it's 2 by 2 metre, 1 by 1 metre, not these big kind of cloud formations, these smaller ones that tend to be particularly difficult to model.

Philip Stier:

Yes. It's almost less the size of the cloud. It's more that clouds and some emergent phenomenon that basically it's a massive multi-scale problem. So clouds are just a collection of droplets and they're formed on aerosols, and aerosols form on the order of nanometer scale and then grow to maybe 2 micrometres and cloud droplets start at that size. But then the droplets are really on the micrometre scale, so very, very small. But then the cloud itself sort of evolves from this collection of droplets and the associated exchange of energy and water. And then on scales of tens of metres, tenths of metres all the way to hundreds of kilometres, but ultimately they affect the global circulation. And when you then go from sort of nanometer scale of the aerosol formation to the global circulation, you have something of the order of 10 to the 14 orders of magnitude scales.

So computationally, we can't resolve all these processes, and that brings us back to these simplified assumptions we have to make, these parameterizations we put in climate models and the inherent uncertainties. And that's the challenge. And that's where we hope AI comes in and will help us to improve the parameterizations. We also constrain them better with observations, but we are just about to open this box.

Jason Mitchell:

Yeah. So let me jump between those two points. Basically, if a global climate model can't explicitly represent everything happening inside a cloud, to what degree can AI help, I guess, bridge that gap? Where would you let a model learn from data and where would you insist that it obey physical laws, particularly around climate models?

Philip Stier:

Well, ultimately the climate model, any climate model needs to obey physical laws. I think where there's currently differences in approaches is we can either constrain this by construction in the setup of the models, and the traditional physically climate model is just a numerical solution of the set of physical equations. Or we can hope to some degree that a model trained on sufficient observational data learns the physics on its own. And these are sort of contrasting approaches. But even if a model has to learn the physics from the data itself, we need to ensure that this is the case and that no conservation laws are violated, that all these constraints are met, and that's a big challenge.

Jason Mitchell:

I want to come back to this idea of what climate models could look like in 10 years. So instead of AI replacing physics-based models, do you expect them to evolve into hybrids? You mentioned this earlier during your first response, but these hybrids that both use physical laws to describe the climate system and learn from data where current models struggle like, let's say, clouds or rainfall. I guess within that, where should AI sit in that picture and what should remain governed by physics and less governed? Is the most pragmatic way to understand in sample extremes using AI and out of sample extremes using these physics-based models? I guess I'm trying to understand the utility function of these hybrid models and what they're good at.

Philip Stier:

Excellent question. I still may not have all the answers. I think my personal view in the future is that we probably will have several types of models available now. So for almost 50 years, we had one type of a physically climate model that resolves the globe. And I think there will be a mixture of approaches. I'm quite certain there will be physically climate models in a traditional sense available, but some of the representations of clouds and other things might be improved by AI. So I think that is one class of models. There will be also high resolution physical models is kilometre scale, sub-kilometre scale resolution models that are really exciting, and I think they will coexist with others. There has been a slightly unhelpful flash of opinions about traditional climate models and high resolution models, and to some degree I think they will coexist for some time.

And now we have pure AI models coming in. They're really very new. We see the first generation coming out. I don't think they solve the problem just yet, but they will coexist and they might help us to tackle different parts of the problem.

Jason Mitchell:

Since the 1950s, better computing observations and models have obviously really transformed weather forecasting. You hinted at this a little bit earlier. In fact, actually the UK's med office says its four-day forecast is now as accurate as its one-day forecast was 30 years ago, which is pretty amazing. AI is obviously bringing big advances in weather prediction. And I'm wondering how much of that success can we reasonably expect to carry over to climate prediction more on a year to decadal timescale? It just feels like there's this big bridge between these physics-based models and these sort of exercise we're doing 30, 50 or 70 years out and these four-day forecasts. How do we fill obviously a very, very wide gap?

Philip Stier:

Yeah, it's a wide gap and there's many different timescales. So as you already hinted, so in weather forecasting, it's already clear within less than a decade AI has transformed the field. And I think that's very easy to verify too, because we can test these forecasts and we can measure their reliability, we can see how well they're calibrated, and we can really say that this has to be a gigantic leap. As we now go to longer and longer timescale, this gets a little bit harder, so there's the transition to the decadal timescale. It's an interesting timescale because on these scales, you still have predictability of the climate system due to slower components of the climate system like the ocean. And I expect there to be a lot of gains still from AI-based prediction system. As we go to the very longer time scales, this might change.

And the difficulty of all of these approaches is we need to ensure that any AI-based component of a climate model, or indeed a purely AI climate model, it needs to generalise to future states. And climate change by definition is an out of distribution problem. We want to predict something that is not in the training data set of these models, and that makes it such a hard challenge. And that is particularly true, as you mentioned, extremes and other events because they're very rare and they may not be well represented in the training data. So just purely data-driven approaches might struggle there. And it brings us again in this combination of physics-based modelling and AI models that might be needed to tackle some of this.

Jason Mitchell:

But where do you see the most benefits, let's say, emerging over the next several years? Is it sort of the near future? And by that, I mean pushing out that four or five-day forecast to beyond a week or two weeks or three weeks. Is there a sweet spot in between in the years to decadal dimension? Is it the far off future sort of backfilling from the assumptions around these long-term physics-based models?

Philip Stier:

Personally, expect there to be quite a transformation on any timescale there, but it will be a bit harder to prove some of this. On the large timescales, it's the same for physically climate models already. Even if you had the perfect climate model that predicts climate perfectly in 2100, you wouldn't be able to show it. So it takes time obviously to get there, and then it takes the development of trust that you would want to rely on such a model.

But then in particular, I think these shorter timescales where we go into up to decadal prediction where there is a lot of reliance still on initial conditions, so the state of now matters for these shorter timescales, I think that there's a real hope that this helps us to improve predictions a lot. And one reason for that, it's that these AI tools can be a very efficient way to estimate the current state of the climate system and to make them predictions from there. And that is already happening. That's related to the field of data simulation. We use heavily in weather forecasting, but I would also expect this to become more and more driving climate forecast on this shorter timescales.

Jason Mitchell:

Interesting. Investors often look at climate projections out to 2050 to even 2100. But to be honest, many of the decisions they make depend on what happens obviously much sooner. Our own climate scientist, Jane Smyth, who you know calls the weeks to months horizon kind of a predictability desert between weather and climate. Again, I keep pressing you on this, but is that where AI can make the biggest difference over the next five to 10 years, for example, in let's say anticipating a monsoon or a growing season or European winter? What could it realistically tell investors about extreme weather risk and what would remain still unknowable?

Philip Stier:

Yeah, you highlight this sort of sweet spot for science, which is also a desert maybe for others. It's a very interesting question because on these timescales, in the atmosphere, you miss predictability. So the atmosphere is a non-linear chaotic system and generally predictability horizon is something like 14 days and then it becomes quite chaotic. But as you're now going into these sub-seasonal to seasonal predictions, then you get information not necessarily from the atmospheric original state, but then from slower components of the climate system. So the land surface becomes important, the ocean surface becomes important, and that makes it quite interesting that you can extend a predictability horizon to some degree.

That's not to say it's easy. There have been very famous failed seasonal predictions, like a famous barbecue summer that never materialised. And it's not a straightforward research question to resolve this, but I personally believe that tools will play a big role exactly in the sweet spot and beyond.

Jason Mitchell:

We've talked about how traditional climate models are built around the physics of the atmosphere and oceans, but when people say AI will transform climate modelling, what do you think will actually change first? Would it be the speed at which we run models, the detail that they can capture, or the scientific questions that we can start to ask?

Philip Stier:

Well, possibly everything really. I think there's a good potential that all of these things will change at the same time. So for these hybrid models, so the combining physics and AI, then speed up is a major advantage. It will be much faster to run, which means you can run larger ensembles, which allow you to estimate the extreme events much better because it's these rare events that you have to capture, so you have to run these models very long. But what is really most important is to reduce the uncertainty of some of these parameterizations, in particular of clouds, because they tend to drive a lot of the uncertainty in climate predictions.

At the same time, if you think of the pure AI models, they will be extremely fast to run, even though they're slow to train. So there's a lot of compute that you have upfront by training these models, but once you have them, the inference is very quick. But they open up the potential for very large assembles for uncertainty quantification, which is very good. But all again, just within model uncertainty, and that is a big problem. You cannot overstress this, that it's an out of distribution problem. And if you train an AI model in present day observations, you may miss events. For example, we are just entering this massive ENSO event that will drive a lot of climate impacts over the next month. It's building up, but would a forecast that has not seen this in the training data be reliable and to predict extremes next year? It's not sure, and this makes it a fascinating question to look into.

Jason Mitchell:

Yeah, I do think it's really important to highlight because I run into this kind of question a lot, but AI is really good at modelling in distribution extremes. These traditional physics-based models are very good. The point is to model these out of distribution extremes for anyone in the audience. But a lot of AI weather models learn from this notion of re-analysis. In other words, historical observations combined with the output of conventional weather models. Does that mean AI is learning something new about the atmosphere or partly learning to reproduce the assumptions and blind spots of those models? And how would we even tell? I'm thinking about the general quality of observational records, which are, let's be honest, they're imperfect in terms of accuracy or coverage, the funding of observational systems of which many are public sector-oriented and underfunded or the gaps of developing countries. And by the way, it's also worth pointing out that Google's neural GCM trains directly on satellite precipitation observations rather than just re-analysis or high-res simulations.

Philip Stier:

Yeah, it's a good question. I don't think we can tell it easily apart what these models learn. If you're trained on re-analysis, you would train on the combination of observations and the physics model. And you inherit some of the bias from either, and that's interesting. For example, even re-analysis doesn't always have the general cycle of precipitation particularly well-represented. It's sometimes related to the representation of convective cloud, and that is something that some of these models will equally learn. So we can't totally take this apart, but it makes it interesting to look into.

So for weather forecasting, this may not really matter too much because as long as your skill improves as a weather forecaster, you're happy and you can verify this quite easily. But for climate models trained on re-analysis, this may be an issue because we have to work out how this affects climate predictions. And then thinking particular about cloud feedback, so the response of clouds to a changing climate, that then could either accelerate or decelerate climate change.

And it's not clear how training on the past 40-year record and observations a model would be able to learn some of these cloud feedbacks. And if it doesn't learn them, it could be potentially underestimate the acceleration of climate that seems to be indicative at the moment by slightly more positive cloud feedbacks.

So you mentioned Neural GCM, so it's a hybrid physics model, so physics AI model. And you rightfully said it early on didn't have reliable precipitation, and they've brought in now training on observation. It's actually a fuse product. It's not a direct measurement of precipitation. It's already a sort of multisensor fusion project they're trained on. It will definitely improve the representation of presentation in present day. But again, the question is will that necessarily generalise for the future?

But personally, I'm really then excited about approaches that come out now where we directly learn the physics from observations without having seen any of the equations. And there's some evidence that these models learned some of the underlying physics like the Navier-Stokes equation of motion of fluids and others just by having seen the data itself. And the only key requirement for this is that the relevant physics is well represented in the training data, which may be often true for the mean, but not necessarily for all the extremes.

Jason Mitchell:

That would be pretty amazing. Can I pull back just a little bit? Because I hear you and Myles Allen talk about the importance of re-analysis, but maybe can you frame that for a second on how big a step function improvement, how transformational is re-analysis for the context of the audience?

Philip Stier:

Well, at the moment it's been the foundation of all of the emerging AI weather models and so on, except as I just mentioned, there is a new generation and that also partly happens at ECMWF, that's the European Centre for Medium Range Weather Forecast that do some of the key re-analysis data that most people use. But at the same time, they're also developing transformative methods to learn purely from data without combining it with a physical model, and then also create a re-analysis from that, but also train models purely for weather forecasting, and they could ultimately become also precursors of climate models.

So I think it's the combination of observations and bring them together in a well-defined physical state is what re-analysis does extremely well, is extremely important. But equally, as we now go to this pure data-driven methods, it is often to have the observations available at scale and for all the purposes that are needed.

I think that's an interesting thing to point out here that in the climate community, we've known for a very long time that some of the key processes we worry about, like cloud feedbacks, like aerosol cloud interactions, that they're driven by these small scale processes and they're not necessarily always well captured by the satellite data alone. That's why we do a lot of research measurements. We do measurement campaigns with aircraft, with drones, with many different things. And at the moment, this data that we think is key to constrain our traditional climate models is not yet something that's used to constrain the AI models because they're mostly focusing on weather. But without this additional information from research data that we create as a community, I think these models will struggle to generalise the future climate. So there's a bit of work to be done.

Jason Mitchell:

Interesting. So we've talked about using AI to make climate models faster or even better at representing physical processes, but a lot of climate science happens around the model itself too, like testing it, analysing the output, and exploring different emission scenarios. Where do you see AI making the biggest difference beyond the simulation itself?

Philip Stier:

Yeah, I think, again, it's a risk of repeating myself, but I think it will be transformative in particular in this aspect as well. So it will allow us to really fully explore the underlying data and make it easily accessible even through text-based interfaces instead of a very complex analysis code that you need some scientists to write for you. And all the ingredients for this are basically there. So we demonstrated a while ago that in a scoping study with Climate Bench in which we trained AI emulators to learn the emission scenarios simulated by a climate model. But then once you have trained this emulator, you can interrogate other emission scenarios that are not trained on, but you can basically interpolate between these emission scenarios. And then you would have a very simple tool to analyse all different emission pathways very efficiently.

On scaling this to the full output of climate model intercomparisons and connecting it then with a large language model allows you to create a text interface for that is totally possible. I'd be very surprised if nobody did this for the next climate modelling exercise because it's just such a nice problem to tackle. But it's a bit of a scaling problem, so it's a huge amount of data, petabytes of data. It's not necessarily best task for academic research groups, but I'm quite certain somebody will do this. And then it will allow a user or an investor or anyone to interrogate the full climate model scenarios simply through a text-based interface.

Jason Mitchell:

I guess where I was getting that was that investors, insurers, and governments often care less about average conditions than about an exceptional flood or a heat wave combined with, let's say, drought. Cheaper AI models let researchers run enormous ensembles in the sense. Does that make these tail risks more measurable or can it create a false sense of precision about events we've rarely or even never observed?

Maybe talk about the difference in value between running a million different plausible weather simulations and knowing the probability of an unprecedented event in a warming climate, that difference between those two.

Philip Stier:

Yeah, I'm really glad you picked this up because it's such an important thing to focus on, is that yes, by having large ensembles, very large ensembles, huge ensembles, whatever people call these in subsequent papers, it does create this false sense of certainty. And often certainties are reported at very high digits, even though it's not clear at all that this is the true uncertainty. When you talk about extreme events, it is exactly these rare tails of the distributions and even AI models, you would need very large ensembles of AI models to actually sample these tails of the distributions. And even in these cheaper models, it's very expensive to do.

So there is a lot of work we and others are working on methods to guide generative models to efficiently sample these rare tails of extreme events and then make this more efficient to get a sampling of extremes. And technically this is promising, and I think there will be lots of progress in this field, but that's only useful as long as these models are really reliable and well calibrated. So if we sample the tail of a model perfectly, but the model has biases that affect the tail, then we still have a massive issue and that could affect all the estimates of extreme events and changing extreme events in the future. So how to do this, I think we are just beginning to find out, but there's a lot of work going on.

Jason Mitchell:

Yeah, I'm sort of mulling this over. And I guess, look, AI tends to be stochastic and will show some reversion to the mean. If we end up, I guess, relying more on non-deterministic models and weather models, will we end up losing information on those extremes? Is that the big risk?

Philip Stier:

Not necessarily, because I think the community recognises this. It depends how you train your models, and when you go to the first weather models that were very smooth because you lost favours very smooth distributions. And people have obviously picked this up, and now with generative modelling, a lot of this is being addressed. And you can bite training in two observations, you can really represent the underlying data structure much better than early on. So I think that there's hope.

So again, because it's quite timely at the moment, we don't have many ENSO event in the training cycle, and we already have strong evidence that ENSO is amplifying and we see this super event just building up, and that wouldn't be represented in the training data. So when you think of tackling some of these extreme events and with purely data-driven approaches, trained on existing observations, you may just simply struggle because the record is so short. For example, when you think of ENSO as with a long, long variability timescale, you only have a few events in the training data set because we have no observations of prior events, and we have no indication if we ever captured such extreme ENSO events as we are now seeing coming up in any of the training data set. Indeed, it's quite likely it's not in the training dataset. So this might really be a case where we see limits of purely AI-driven climate forecasts here and where the combination of AI with physics-based models will be really crucial.

Jason Mitchell:

Got it. I'm going to jump lanes a little bit here, but what do you see as the role of private companies in this space? Anxiety about the climate data, call it the industrial complex, have been around for more than a decade, but this time seems a little bit different with companies like Google, Microsoft, Nvidia now developing and funding many of these tools. Beyond raw computing power, what do they contribute that universities and public weather services can't really? What's at risk becoming vulnerable if essential forecasting capability depends on a smaller and a smaller number of private companies?

Philip Stier:

Yeah, I mean personally, I believe this is extremely welcome friendly competition, so it's exciting to see this work and the progress in particular in the weather space has been transformative. It's really pushing everyone a bit, so that's good. These companies have amazing resources. They have big pocket-size, some of the smartest people and trained all of by us in academia. And all of this is good. It provides a fantastic career path for people leaving academia and going into these companies. So that's great.

I think at this point it's not entirely clear yet where this is leading and what some of the underlying drivers for the companies are. So for weather, there's a very clear and strong business case. So that's clear that where some of the tech industry will want to go with it, and having the perfect weather forecast in your pocket on the phone is obviously hugely important to many. And that's not just the daily weather forecast, it's extreme event warnings and many things. So maybe on the daily or sub-daily to seasonal time scales, I think it's a strong business case to be made.

For climate itself, it's a bit less obvious and I'm curious to see how this evolves. So climate model development often involves failures. You may develop a model and it may not be looking as good as others. And that in academia and in the public research sector, climate research sector, we're used to it, but totally fine with it. But I think in tech, this is a bit more frowned open, not to be on top of the leadership board. And so I'm curious to see how many people still want to enter the leadership board if you're not entirely sure where you are.

So I think it's a very rapidly evolving space. But now coming to your part of the question about reliance on the private sector, and I think I agree with you, climate change is one of the biggest challenges to humankind, and it's a central public good, we need to understand this, we need to make good predictions, we need to be able to warn people. And we cannot afford to give away the capability to make climate predictions to the private sector alone so where priorities can change extremely fast. So there's definitely a very strong need for strong public funded and open sourced climate research as it's happening already now, but maybe not on the scale that it should be funded at, to ensure that this is a long-term or that the tools are long-term available for greater good.

Jason Mitchell:

I guess within that kind of discussion, what should open really mean? If a company publishes its climate model, but others can't examine its training data, reproduce its result or test it on the events they care about, is that really open science? It kind of feels like replication studies could have a real purpose here. So what minimum standard would you want for a model used in public climate decisions?

Philip Stier:

Good question. It's a concern. And I think open... I mean it was interesting in traditional climate model development, we've gone from closed to very open over the last 20, 30 years. Most of these codes were entirely closed and not really openly accessible. Now if you develop and publish it, you basically have to open source your code and everyone can interrogate it and typically you can just also run anyone's model. So that's great. In many of the AI weather models, the code is open to some degree that you can have the model and the weight, you can maybe run it if you can create the initial conditions, which actually very few people can create because it's again, this data simulation problem combining the observations in a good way. So that is open, but at the moment the public cannot recreate these models easily. We cannot easily follow this training pipeline. We don't have access to the data, we don't have access to some of the tools, the software, but we don't even have access to some of the hardware use in some of these training pipelines. So the public sector couldn't do that, and that is sort of a limit to openness.

So replication studies or benchmarks can get you somewhere in particular with AI-based models, but there's a limit because you can, in these models with many, many parameters, hundreds of million, billion parameters, you can fit an elephant. So overfitting is a real risk. And these models could very easily look perfect on all past observations because they fit very well to them, but that's no indication of their generalisation to future climates, and that's a challenge.

But then ultimately, it comes back to trust. And trust in climate modelling is something that's hard earned. So in traditional climate models, we trust them because we know the underlying physics, we know they encode the underlying physics. Where the uncertainty comes in is then how you represent this multi-scale physics problem. But fundamentally, there's a trust in the physics while when we come to AI-based models, it's not entirely clear where it would arise from because fitting to past observations and showing you perform well in it does not necessarily give you trust you generalise in a future climate.

Jason Mitchell:

Interesting. Interesting. To the degree you've had exposure to investors, in what ways do you think that climate information can be delivered more effectively to inform financial and other economic decisions? I guess there's this sort of outstanding question of are GCMs and ensembles fit for purpose in a financial and economic context. It almost seems like a key bottleneck, not only in pricing climate risk, but also making climate information more actionable, let's say, in a real world context.

Philip Stier:

Yeah, I've been lucky enough to have many fantastic discussions with investors over the last years in the financial sector in general with you and others. And it was really interesting to see how the importance of climate is being more and more recognised. So I think that's a fantastic start from. There's also, I feel, very strongly growing expertise in the sector. So many financial institutions have teams dedicated to the problem because it's such a massive problem. It's a very sensible thing to do. So things definitely in that sense go in the right direction.

As for delivery of information, I think a lot can always be improved. Well, that's true in general, but also in this sector. I think while there's a really good general understanding of climate models as such and the output, I think the understanding of uncertainty is sometimes less developed and that is a concern. There's often reliance on these climate model intercomparisons that we use, this CMIP, Climate Model Intercomparison Project output, which is a very important data set and certainly provides a clear, very relevant background for decision making.

But if you're in the climate modelling community yourself or in the climate physics community, you're also much more aware of the limitations just because a CMIP model predicts something will happen in your backyard because it's the nearest grid point in the model, doesn't necessarily reflect the risk of changing climate risk in your backyard in these very cost resolution models. So I think it needs this sort of interaction.

But in terms for the sector, what will change quite a bit is that these AI-based climate information systems we touched on before, the emulation of climate model output and how you will be able to interrogate them without expert knowledge. Most likely just with text-based interfaces, you can get any climate information on the right scale out that will transform the field. There may be a few years out, and I think the real challenge will be in these systems, again, how you communicate uncertainty, because we touched on before you can spit out any uncertainty with any... You can give many numbers on how accurate the uncertainty might be, but you need to understand really the underlying systemic uncertainty in the models. And that's a challenge. It's a few years out, but I think on the 10-year horizon we are talking about, yeah, I think it will change everything.

Jason Mitchell:

Last question. Along these lines, where do you see blind spots in terms of finance ignoring or overlooking climate phenomena? What do you think are the most undervalued or underpriced physical risks in current financial markets based on your modelling experience? Are there specific geographies or hazards and let's say wet bulb heat or is it water stress or wildfire weather or other phenomena that deserve more attention from investors?

Philip Stier:

A good question that I'm not nearly an expert in this view really, but more generally I think it is coming back to this blind reliance on current climate model output. It can be an issue where the uncertainty isn't necessarily appreciated. More specifically, when we talk about predictions of extreme events from precipitation and wind, it's often what we call organised convections. So it's not just a single cloud, but clouds itself organises as a larger self-organised system and it's responsible for significant damage worldwide. It goes all the way, obviously tropical cyclones and hurricanes are a very big organised connective system, but it's still the same class of systems. And these are just not well represented in current low resolution climate models.

And so if you just rely on some of these outputs, you may just underappreciate and undervalue this risk. There's obviously methods that people use to bias correct for this and use other tools to take into account. But I think at the moment, climate modelling as such is just not there to deliver some of this. There's a lot of interesting work going into it, but for some of these organisations, you need to represent these systems more explicitly than we currently do. And even hybrid climate models, as long as they run on low resolution, they will struggle to explicitly represent this organisation of convection and that drives a lot of the damage worldwide.

So that's, I think, one of these, but then there's many other factors that we need to take into account. There's, as you say, heat stress, water stress, wildfires, but then particularly from recent sad example, glacier stability. We currently do not have a prediction system that would tell us of the glacier stability now on a seasonal timescale where people might have early warning systems for glacier collapse. So there's a lot we need to do. There's a lot that once we as a climate science community develop it, that we need to translate back to users and the finance sectors and other sectors. The question is really how to do this best. But to some degree, I think close collaboration with the academic sector and the users will be absolutely key.

Jason Mitchell:

Great, great. Good way to end. So it's been fascinating to talk about what AI could transform in climate modelling, where it can genuinely extend what physics and earth observations currently tell us, and ultimately how better climate information can lead to better decisions and better pricing of risk. So I'd really like to thank you for your time and insights.

I'm Jason Mitchell, CIO for Responsible Investment at Man Group, here today with Professor Philip Stier at the University of Oxford and director of its Intelligent Earth Centre. Many thanks for joining us on A Sustainable Future. And I hope you'll join us on our next podcast episode.

Philip, thank you so much for your time today. This has been super, super interesting.

Philip Stier:

Fantastic. It was great to chat to you.