Episode Categories:
Resources:
Join the discussion: Ask Jordan and Mike questions on AI and weather
Upcoming Episodes: Check out the upcoming Everyday AI Livestream lineup
Connect with Jordan Wilson: LinkedIn Profile
Try Our Free AI Prompting Course: Register for our free Prime, Prompt, and Polish AI Course!
The Role of AI in Weather Forecasts and Climate Modeling
As technology evolves, artificial intelligence (AI) continues to advance sectors not previously associated with it. A notable example is weather forecasting and climate modeling. Meteorological predictions necessitate a blend of data sourced from satellites, aircraft sensors, and ground-based sensors, generating a representation of atmospheric conditions. AI enhances the accuracy of these predictions by producing probabilistic outcomes, essentially managing the stochastic elements crucial for high-resolution weather modeling.
Private Involvement in Weather Prediction: NVIDIA’s Contribution
Traditionally, government-operated models mastered weather prediction. However, AI models honed by private enterprises are increasingly reaching the skill level of these long-established models. NVIDIA, a pioneer in this enterprise, has made significant strides in enhancing weather prediction technology. Despite not maintaining a proprietary weather prediction model, NVIDIA actively improves existing models through its cutting-edge technology.
Raising the Bar: Impact of NVIDIA’s Stormcast Project
Stormcast, a research initiative by NVIDIA, ups the ante in AI weather prediction models. The project aims to create a high-resolution, national U.S. weather prediction model, exhibiting remarkable accuracy that rivals existing U.S. national weather models despite being in the prototype phase. Such developments indicate how AI could significantly augment climate simulation by bypassing computational limits, paving the way for efficient, high-resolution models.
The Power of Generative AI
Generative AI elucidates the future capabilities of probabilistic prediction by generating multiple plausible scenarios for future conditions—an important feature considering the inherently unpredictable nature of weather systems. For instance, AI weather models possess the potential to operate at a superior speed, possibly a thousand times faster than traditional physics-based models.
Transforming Climate Science Via AI
Through AI, climate science is making significant strides. By surmounting conventional computational barricades, AI enables more detailed, resource-efficient climate simulations. Additionally, AI's promising role in understanding and predicting climate-related phenomena, such as cloud formation, could potentially revolutionize how we address adverse climate effects.
Future Implications
As AI seamlessly integrates into major meteorological processes, the future of weather and climate prediction will heavily involve AI technology. Collaborative efforts across private and public sectors are leading this evolution, with the applications of AI in weather prediction seen as a significant boon for humankind.
Summary
AI's role in transforming weather prediction, emphasized by advancements from NVIDIA in data modeling, signals a paradigm shift in climate science. Soon, the technology might enable on-demand weather forecasts, ushering in an era of personalized weather prediction models. At the heart of these upgrades is more accurate and faster predictions, improving planning, adaptation strategies, and safety protocols. It's clear—the impact of AI and advanced technology in meteorology and climate science is just beginning.
Topics Covered in This Episode
1. Role of NVIDIA in Weather Forecasting
2. Impact of AI on Weather Forecasting
3. Data Streams and Predictions
4. AI in Climate Change and Regulation
Podcast Transcript
Jordan Wilson [00:00:18]:
The weather and climate is something that impacts all of our lives. Especially here recently in the US, we've had a few devastating storms that have really caused a lot of havoc. So can generative AI help us better prepare and predict, for these storms, and how can it improve our climate in the long run? Well, that's what we're gonna be talking about today on everyday AI. I'm I'm very excited to have a leading expert from NVIDIA to help us go over all of this very timely and very important topic. Alright. Well, what's going on y'all? My name is Jordan Wilson, and welcome to Everyday AI. Before we get started, I have to give a shout out to our partners at Microsoft. So the WorkLab podcast from Microsoft is made for leaders who want to understand the future of work.
Jordan Wilson [00:01:13]:
It offers expert insights on everything from how to approach digital transformation to what it takes to thrive in the AI area. That's Worklab. No spaces available wherever you get your podcast. Speaking of podcast, well, you're listening to it right now. Welcome to Everyday AI. This is for you. It's for me. It's for all of us to help us better understand generative AI.
Jordan Wilson [00:01:35]:
So, if you haven't already, if you're listening whether on the podcast, the live stream, or reading about this on the newsletter, well, please go read that newsletter. Go to your everyday ai.com and sign up for that free daily newsletter. Alright. So I am super excited to talk about how AI and generative AI is impacting the weather and climate. But first, let's go over quickly, started how we do every single day by going over the AI news. So Microsoft has introduced new AI tools to help revolutionize health care efficiency. So Microsoft has unveiled a series of advanced health care data and artificial intelligence tools at improving workflows for clinicians and reducing burnout in the industry. So the new offerings include a collection of open source multimodal AI models that can analyze diverse data types, such as medical images, clinical records, and genomic data, allowing health care organizations to develop tailored applications more efficiently.
Jordan Wilson [00:02:31]:
So Microsoft has indicated that many of those tools are still in early development stages or available for preview testing, emphasizing the importance of validation by health care organizations before wider rollout. Hey. Speaking of NVIDIA, NVIDIA had their stock is surging after black their new Blackwell GPU has reportedly sold out for a year. So NVIDIA's stock is experiencing a significant uptick following the announcement that demand for its upcoming Blackwell GPU has exceeded supply resulting in a complete sellout for the next months according to reports. So this surge in interest obviously reflects the growing appetite in advanced GPU demand in various sectors. So, this announcement has led to a noticeable increase in NVIDIA stock price where it is near an all time high. And keep an eye out whether today or to, today or early next week, where in hey. NVIDIA might be passing Apple, to become the most valuable company in terms of market cap to reclaim that title.
Jordan Wilson [00:03:35]:
We always cover who's who's the most valuable, you know, company in the world, here on the show. So, our last piece of AI news, Tesla has unveiled a new line of AI powered vehicles, last night at their We Robots event. So they announced the the cyber cab, and the cyber cab is designed to operate entirely without human intervention, showcasing Tesla's advancements in AI and self driving capabilities. So this vehicle aims to create a new standard for ride sharing services. So Elon Musk revealed the cyber cab and that it will cost reportedly below $30,000 and is expected to launch by 2026. So this price point could make autonomous transportation more accessible to a broader audience, potentially transforming how people commute. They also announced the robo van, a larger autonomous vehicle capable of seating 20 people, which can be adopted for various uses like a school bus, an RV, or for cargo. Alright.
Jordan Wilson [00:04:31]:
There's a lot more, where that came from, but make sure to go check out the, newsletter for that. Alright. I'm super excited for our guest today. I think the topic of weather and climate is very important for us all to pay attention to. Not only that, but how can generative AI actually change it? So, I'm very excited to have him on the show. Please help me in welcoming, there we go. We got him, Mike Pritchard, and Mike is the director of climate simulation research at NVIDIA. Mike, thank you so much for joining the Everyday AI Show.
Mike Pitchard [00:05:04]:
Thanks so much for having me.
Jordan Wilson [00:05:06]:
Alright. Hey. And we gotta shout out those of us that that, you know, join at, like, 5:30 AM West Coast time. So, you you know, Mike, thank you for that. So so real quick, kind of before we dive deeper into this topic, can you tell us a little bit about what your role entails there at NVIDIA? Sure.
Mike Pitchard [00:05:23]:
Yeah. I'm I'm the director of climate simulation research. It's my great privilege to have been, brought over from academia to NVIDIA and to lead a team of 4 or 5 AI researchers and dead climate domain experts who are working as part of a broader initiative called Earth 2 to try to, do fundamental research where AI has the potential to transform the Earth system modeling stack from weather to climate, from near term to future.
Jordan Wilson [00:05:46]:
Yeah. And I was at, you know, partnered with NVIDIA for the GTC conference, and I was on the floor when, you you know, they put Earth 2 up on the screen, and I was, like, mind blown by that. But, you know, maybe can you just explain, Mike a little bit because I think a lot of times people think NVIDIA. Right? And they think maybe traditional, like, traditionally gaming. Right? GPUs for gaming and now, you know, everyone knows, NVIDIA for, you know, providing those chips that power everyone else's generative AI systems. But how does how did NVIDIA kind of get involved in in weather and climate? And, you know, what is really your goal in that space?
Mike Pitchard [00:06:24]:
Oh, what a great question. Yeah. You know, I had the same impression. I'm I'm an academic. I've been a professor of climate science since 2013, and my my first experience of NVIDIA was through using their supercomputers, which which changed my life. You know, a GPU powered supercomputer that the Department of Energy, bought eventually allowed me to do climate simulations a 100 times more ambitious than I was used to. But what I've come to discover since actually joining the company is there's a a research organization that just does fundamental AI research on on cutting edge generative AI, including on some applications that are important to the company. And and it's it's Justin Huang, our CAO, that's declared climate as something very important to the company.
Mike Pitchard [00:07:00]:
He actually gave a great keynote at the Berlin Climate Summit last July. I encourage anyone to watch if they're interested in our positioning. There's 3 fundamental missions of Earth 2. The first is to accelerate global cloud resolving modeling technology so that we can perform higher resolution simulations of the Earth's climate than has ever been possible, and using AI to enable more interactive experiences of the output, and then also to produce stunning visualizations. And I guess you saw one at the GTC conference in March. I hope that gives you a taste. Oh, I should also say NVIDIA's got a long history of fundamentally accelerating weather and climate of coast that goes back at least a decade. And and I've been amazed to discover how how NVIDIA engineers have been inside many of the, the code bases I've was really with in academia.
Jordan Wilson [00:07:44]:
Yeah. And, you you know, I'm sure there's gonna be people, tuning in today who have experience, in this field. But I think for the most of us, when we think weather, right, we just, you know, turn on the weather channel or maybe maybe that's aging me. Right? Like, I always like listening to my local meteorologist here in Chicago or, you know, opening up apps. But maybe, Mike, can you just talk in general? Right? So even, you know, outside of NVIDIA or or not, but how does the weather actually work before we can actually understand how generative AI and all the research that you're doing impacts it? Like, how do, like, how do we get all this information, and and and then we'll go, a a little bit into how NVIDIA's changing that.
Mike Pitchard [00:08:25]:
Sure. Yeah. Great. Yeah. Yeah. Behind your phone app, there's an incredible enterprise going on. You know, every, it begins with observations. At every 6 hours, there's a coordinated launch of balloons worldwide that sample the atmosphere.
Mike Pitchard [00:08:35]:
And these data streams, along with satellites and aircraft sensors and ground based sensors, are fused into a best estimate of the initial condition of the atmosphere that's used to produce predictions, to watch predictions from. And the predictions are done by physics models, you know, computational models that solve equations like Newton's law, f equals m a on a on a rotating sphere for an incompressible gas. And and and these predictions into the future, they're chaotic. They're stochastic because of the butterfly effect. So you have to do many predictions of 6 hours from now because things could turn out differently. But but it's a very computationally intensive exercise and, you know, your average meteorologist is looking at the results from, you know, a half dozen to a dozen different national and international weather models and making their own nuanced decision about what to tell the public based on their familiarity with those models and their biopsies and their behaviors and, and their observations from the sensor network of what's going on. So there's a lot going on behind the scenes of your phone.
Jordan Wilson [00:09:33]:
Okay. That's that's super helpful. So essentially, you know, all the meteorologists and, you you know, these these organizations that help us better understand the current weather, they're looking at many different models. I guess is does NVIDIA have one of those models that everyone else looks at, or is maybe NVIDIA's technology helping all of those other kind of companies, improve their respective models. I hope that makes sense. But, you know, I I love weather and, you know, I'm I'm I'm trying to really, explain it here for everyone else to see how NVIDIA's work is currently, you you know, helping and how it will change.
Mike Pitchard [00:10:06]:
That's a great question. Yeah. Definitely the latter. You know, in the research org, we're trying to seed technologies that we hope will will will transform the business of weather prediction that will improve it. And and we're already seeing some some uptake. It's fascinating. You know? I'm a climate scientist, so I've mostly been concerned about 2020 to a 100 years from now. But when I joined NVIDIA is when I discovered what was going on with the AI revolution in weather.
Mike Pitchard [00:10:28]:
And, you know, NVIDIA is proud to have built one of the first global AI weather models that could feed off the the native resolution of one of the highest quality datasets, worth training on. And, I think, it's amazing to look back 2 years later. This seed technology has now disrupted the world's premier weather prediction agency, the European Center For Medium Range Weather Forecasting, who has developed an AI companion model for their classical physics model that's starting to beat their classical physics model and is being served to the public alongside their their standard physical predictions. So this is happening all around the world in the private sector, in the public center. Large meteorological agencies are taking note of the potential of AI to beat physics, predicting the weather.
Jordan Wilson [00:11:11]:
Wow. And and that's to me, that's wild. Right? But, you know, can you even maybe explain specifically, how does generative AI change this? Right? Because NVIDIA is, one of the companies leading the generative AI revolution in many different ways, not just from a hardware perspective, but, you know, the infrastructure that you help, provide other companies as as well. But how does generative AI change? What can be done, with, you know, climate, you know, climate prediction, weather? How does generative AI change at all?
Mike Pitchard [00:11:45]:
Oh, thank you so much. You're gonna have to keep me on the rails short of me because I nerd out too hard here
Jordan Wilson [00:11:50]:
about Hey. That's that's what this is for. We're, like, we're here to nerd out a little bit. Right?
Mike Pitchard [00:11:54]:
Good. Okay. To to me as a physicist, generative AI is so important because it can produce stochastic predictions. It can produce an ensemble of probabilistic outcomes. And, you know, so so for the topic of weather prediction, we're limited by the data. And you mentioned Stormcast. This is our research paper I'm so excited about. A challenge we faced in building Stormcast, which is, like, trying to become a high resolution national US weather model, is that the training data is available only at low frequency and time, so about an hour.
Mike Pitchard [00:12:26]:
That might seem high frequency, but your average storm, you know, an individual cumulus cloud lives for only an hour or 2. So that's not a lot of data to understand the process that's going on inside the storm. And especially if you're predicting those high resolution physics on an hour time scale, there's many possible outcomes of the next hour ahead. And so generative AI is a way to to respect the fact that the physics are stochastic, that the prediction is probabilistic, and to produce an ensemble of equivalently plausible outcomes for an hour ahead. And that's that's deeply important to the business of weather and climate prediction, this ability to do probabilistic prediction.
Jordan Wilson [00:13:03]:
Okay. So you've you you you've kind of mentioned storm cast a little bit. Can you explain this piece of of research by NVIDIA? I'm personally fascinated by it. So, yeah, can you explain to us what what is a a storm cast?
Mike Pitchard [00:13:16]:
Okay. Thanks. Yeah. So so if you I mentioned that, there's many different meteorological models out there in the world that are served to you, and there's different flavors of them. So there are planetary scale models that cover the whole planet with 25 kilometer resolution, you know, about the scale of a county. And then there's national models, countries that can afford it, like the US, have high resolution national models where the resolution is on the order of a kilometer. And, it's it's too computationally expensive to go to the whole planet with that much resolution, but but countries cover their their own domains with that much resolution. And, for 2 or 3 years now, these global models, these course resolution models, AI has completely disrupted them.
Mike Pitchard [00:13:56]:
And I've been waiting for this moment when AI might also disrupt the business of these high resolution models where the physics are are quite different, the physics of thunderstorms, the physics of cloud formation, the physics of turbulence. And and so that's what Stormcast is. It's a a solution to trying to build an emulator, an AI emulator of the US national weather model. As a proof of concept, we only study a patch of atmosphere in the central US, 900 by 900 kilometers, but it performs predictions in real time that you can compare the skill of these predictions against the US national weather model. And it's got as much, if not sometimes more skill, but despite being a prototype in its infancy, as the National Oceanographic and Atmospheric Administration's high resolution rapid refresh model, which is what's served to you if you use windy, for instance, on your phone to get your aviation outlook. So that's quite exciting. It's a milestone in AI, at this high resolution prediction
Jordan Wilson [00:14:52]:
scale. Can you because I'm still trying to wrap my my brain around this. Right? How, right, how can, you know, this piece of research, from, you know, a private company already be comparable, to these models that have been used, I think, for decades and, you know, that are, overseen by the government? Like, how is that even possible, and and what does that also mean, for the future of of weather prediction?
Mike Pitchard [00:15:22]:
Well, those are deep questions. I'm sure I won't have the answer to all of them. It's important to recognize that a big part of how is the the huge investment governments have made in producing very high quality, quality controlled, datasets that are in the public domain. And so we should all be grateful that in the US, you know, our our national agencies like NASA and NOAA, are have a mission to serve high quality data to the public in in the open domain. And and so that's essential. I mean, the training data everything flows from the training data. But a big part of the how I've realized, especially coming from academia and my former life, are the personnel. I've never met people like I've met at NVIDIA.
Mike Pitchard [00:15:58]:
You know, professional deep engineers, professional JNI researchers, the peep the sort of people that can take a massive dataset and unplug all the bottlenecks that can prevent you from training efficiently on it at scale and mining its essence. And, but it turns out it is remarkable. You're right to point it out. You know, teams of 5 or 6 people. And in a few separate companies, Huawei and Google and DeepMind, NVIDIA, have managed to produce convincing AI weather simulations that have competitive skill with these huge government institutions. And, yeah, so it it says something deep and, you know, gosh, I can't wait to see what it's gonna look like in 5 years when once the government institutions with all of their domain expertise and history and institutional knowledge of the available data streams also embrace this technology. It's beginning to happen.
Jordan Wilson [00:16:45]:
Wow. Okay. So I I wanna dive in, a little bit deeper on that, but just real quick, we have to give a a quick shout out, to our partners at Microsoft. So the WorkLab podcast from Microsoft is made for leaders who want to understand how work is changing. Effective leaders adapt. They stay on top of trends. They embrace any edge that they can get. Effective leaders also know that the key to understanding artificial intelligence is to get better at understanding human intelligence.
Jordan Wilson [00:17:13]:
So for real world lessons and actionable insights to help you stay ahead, check out the WorkLab podcast. That's WorkLab, no spaces, available wherever you get your podcasts. Alright. So thank you to, our partners there at WorkLab. So let me just get right back to this. So you said, Mike, you you kind of said, hey. Can't wait to see what's going to happen in 5 years. Well, what does that mean? Right? So I I know for people, you know, like you, researchers, people who are working in and around the field, I'm sure it means one thing.
Jordan Wilson [00:17:43]:
What does this mean for everyone else? Right? Like, is is every single you know, you mentioned, you know, meteorologists, they use all these models. Well, essentially, the storm cast research just help all these models become so much better, or or how does this actually impact, you you know, the average person when it comes to weather and climate?
Mike Pitchard [00:18:04]:
Oh, wow. Yeah. There's a there's a few future. I mean, we're daydreaming now here if we're entitled to daydream. There's a few scenarios that Yeah.
Jordan Wilson [00:18:10]:
That's that's what we're here for.
Mike Pitchard [00:18:11]:
Yeah. But I mean, I mean, what one possibility is that AI so disrupts the the business of data assimilation and the initialization of weather models and and all the coordinated mechanisms I mentioned, that weather forecasting becomes more on demand that, you know, users can launch their own forecasts at any time they'd like rather than waiting for the institutions to launch their forecast on the hour or every 6 hours. Another possibility is that there there becomes a a biodiverse universe of weather models that are niche. They're fine tuned to individual stakeholders. You know? People care about the weather for different reasons and in different regions, and weather models are are not necessarily calibrated to individual stakeholders' needs. So there could become a a a a large a larger number of weather models that have ever existed before if the technology becomes fine tunable and differentiated the way, some large language models have for instance. And, but, yeah, I think mostly the biggest thing though is the speed. So once trained, these AI weather models are a 1000 times faster than the physics calculations, and that changes everything.
Mike Pitchard [00:19:16]:
You know? For the history of weather prediction, there's been this tension between how much resolution can I put into my weather forecast? I'd like more because hurricane forecasts are more realistic when there's more resolution, but versus how many ensemble members can I simulate? I'd love to simulate a 1,000 realizations of the hurricane to know what the worst possible one is to hope for the best and plan for the worst. But those things are intention because they trade off against a computational budget. And when you unfollow the compute by a 1,000 with AI, then you can afford massive ensembles. Mhmm. And, you know, I I mentioned the Stormcast technology for the regional scale and the high resolution is new. But what's more mature now are the global AI emulators for the world. And and I'm really there's another couple of research papers that we've done with the with UC Berkeley that I'm really proud of that's proved this massive ensemble capability where you can do, you know, 20,000 realizations of last summer, 2023, to get counterfactuals on a warm summer to understand the, you know, the statistical drivers of low likelihood, high impact extreme events, which are so important to calibrating our risk, to extreme events and to understand, the hazard of future events by knowing what our exposure is today. So these are some of the ways that I think just the the compute, can be transformative.
Jordan Wilson [00:20:35]:
Yeah. So so, Mike, at the very beginning of that answer there, and I I I wanna make sure that I fully understand this and our audience does too. So, you know, you said that kind of the the research that NVIDIA has done with Stormcast, could enable, you know, weather forecasting to become much more on demand. So you said right now, these forecasts, you know, that all meteorologists and everyone else use come out every hour. Right? And now it could be on demand or people could fine tune, these, you know, future models to, you know, be more personalized for them? Is is that correct? And then if so, right, how does that actually change? Right? You know, I kind of started the show by by talking about some of this, you know, devastating, weather that we've seen, recently with with the hurricanes and the tropical storms, especially in the, you know, southern US here. How does that change things in the future? Is it just much more accurate and faster, weather predictions?
Mike Pitchard [00:21:33]:
Great questions. Yes. So I think we're all hoping that eventually there'll be more accuracy for extreme events like hurricanes, you know. There's already a glimmer of hope that the a the global course resolution AI emulators are producing more skillful tropical cyclone tracks than conventional physics calculations, so that's quite encouraging. The the holy grail is the intensification. I mean, we've seen just, like, how perilously the latest hurricane intensified over the over the warm Gulf of Mexico. So that that's a ways off, and there's a lot of research to do. But, to the extent that, you know, in physics calculations, we have to cope with a lot of problems of how we formulate the grid and how we represent unresolved processes that can lead to intensification.
Mike Pitchard [00:22:14]:
The ability of AI to be completely agnostic to that and just feed off data and nonetheless be skillful is is just a philosophical reason to hope that we might penetrate the skill barrier in intensity forecasting. And that's just so important to everybody that wants to plan for an incoming hurricane to get an accurate, reliable intensity forecast. In terms of the the on demand nature, now again, we were daydreaming and painting into the future here. You know, Stormcast has proved the prediction problem, can be tackled with AI that we can get competitive storm, skill, convection skill compared to the national weather model. The initialization problem is a different beast, but it's the marriage of the 2. And there's people working on this, and we're working on it in NVIDIA too, trying to use AI to, to make the process of how you initialize these weather models more seamless in addition to changing the prediction from physics to data.
Jordan Wilson [00:23:05]:
Yeah. And it's it's it's fascinating to to hear about, Mike, and to have you explain it in such a simple way. Me you know, my my brain's churning, and I'm sure everyone else's is as well. You know, even if if if we zoom out. Right? If we talk a little bit more broadly, about the the climate. Right? So you you helped us better understand, you you know, storm cast and how that can help, with more real time and accurate weather predictions. But what if we zoom out just to the climate in general? How is generative AI and the work that you're doing at NVIDIA, helping us look at climate a little differently and what does that mean for everyone?
Mike Pitchard [00:23:43]:
Well, thank you so much. Yes. Yeah. So I'm I'm a climate scientist. I've spent my my career trying to investigate algorithms with breakthrough potential for climate simulation because it it's a horsepower limited problem. Simulating the whole planet for a 100 years, hundreds of times to look across what if scenarios of what emissions might do, that's just a gargantuan computational problem, and you have to make compromises that feel unsatisfying in the number of processes you're resolving. And so AI could short circuit Moore's law. It could you know, when when you double the resolution of an atmospheric model, Jordan, the computational intensity goes up by 8 because you have to double, you know, multiple dimensions of space, and then also the time step has to shrink to keep the simulation stable.
Mike Pitchard [00:24:26]:
So getting high resolution simulations is actually really, really difficult. And if if AI if you can just outsource the high resolution physics to AI, it changes everything. It means that you can you can afford high resolution physics today. So I worry about clouds a lot when it comes to climate. I live in Southern California. I drive by the stratocumulus cloud. We call it the marine layer. It looks like a little gray haze on the horizon when you're at the beach.
Mike Pitchard [00:24:49]:
If it comes in, you're you're annoyed because it makes you cold. But it's just the edge of a massive mirror, of clouds. So you can see them out your window if you fly from San Diego to Hawaii for half the trip, because of massive sheet of clouds, they reflect a lot of watts from the planet that keeps the planet cooler than it would be otherwise, like ice sheets. And if those clouds shrink like ice sheets and dissipate, revealing darker, more absorptive surfaces, it amplifies global warming. It it amplifies all the hazards that go with global warming. If they thicken up and become brighter, which they might, unlike ice sheets, they could they could counteract the hazard. We actually don't know the answer to that question with high confidence because high resolution is the key to simulating those clouds clouds faithfully. So so, the same, you know, the the same business of AI short circuiting Moore's Law, removing tensions that have existed computationally in weather and climate simulation is the reason to be excited for the climate problem.
Mike Pitchard [00:25:41]:
Now I didn't mention generative AI there. Where generative AI is proving very helpful in work with NVIDIA for climate is for downscaling, and that's for turning low resolution predictions, which we have a lot of today from the world's climate, climate modeling agencies into high resolution actionable predictions. So you can just, like, super resolution in images, you can take these low resolution data and turn them into high resolution, impacts, impacts relevant metrics.
Jordan Wilson [00:26:13]:
The WorkLab podcast from Microsoft is made for leaders who want to understand how work is changing. Effective leaders adapt. They stay on top of trends. They embrace any edge that they can get. Effective leaders also know that the key to understanding artificial intelligence is to get better at understanding human intelligence. So for real world lessons and actionable insights to help you stay ahead, check out the WorkLab podcast. That's WorkLab, no spaces, available wherever you get your podcast. So, you know, Mike, we've we we we've gone over, you know, so much here.
Jordan Wilson [00:26:52]:
So, you know, as an example, you mentioned in the beginning, you know, earth 2, which, you know, I think is just kind of like a digital twin of the earth. Right? We talked about storm cast. You you you know, you talked about all this great research that that you and your colleagues are doing, that can help us better understand, you know, earth's climate. You know, all of these things happening at once and then throw in obviously generative AI and, you know, all of the compute resources that you have at NVIDIA. What is your mind focused on? Right? Like like, what are you thinking? Okay. Hey. Once we, you know, figure this piece out, then this happens next. Right? Like, what is that next thing that you're looking at when it comes to this, you know, this, intersection of, you you know, weather, climate, and AI?
Mike Pitchard [00:27:39]:
I'm afraid there's no one thing, Jordan, because there's a few fronts of work that that we need to progress. But on the weather side, I think the frontiers are sub seasonal to seasonal prediction. Evolving these AI weather models that have proven to be good at atmospheric prediction to also be to be good at ocean prediction and coupled ocean atmosphere prediction. The ocean moves slower than the atmosphere. So when you want the predictions to roll up beyond the time scales of weather into climate, beyond weeks to months to years, then you need to start worrying about simulating the ocean with AI, simulating to couple the directions between the ocean and atmosphere with AI, and then especially simulating the whole system's response to changes that humans are making, like turning up the c o two and changing the opacity of the atmosphere or changing land use or changing the emissions of soot particles that can interact with clouds and change their brightness. So each of those things I mentioned are frontiers where, researchers are working actively within NVIDIA, across NVIDIA and other institutions with their collaborators to try to see can the same gains that are being made for the atmosphere, for the weather time scale, really fundamentally disrupt the climate at time scale where these couple interactions become really important. But there's a very complicated physics and multiscale physics that's gonna take a lot of lot of work to see if it it can work if if it can happen. Yeah.
Mike Pitchard [00:28:51]:
And then I think the other the other front that I'm really passionate about though is downscaling. You know, there's so much impact to be had from taking all the wonderful predictions that we already have from the world's investment in physics based climate making them relatable, making them high enough resolution for us to understand their implications for the average person to interact with them, and to just like AI weather can create massive ensembles of weather, to create massive ensembles of these climate projections so that we can hope for the best, but plan for the worst and understand how extreme events, you know, including the most extreme events, may change in the future to to build resilient infrastructure. Mhmm. So those are the two fronts that I'm I'm most passionate about.
Jordan Wilson [00:29:31]:
So, you know, Mike, I know this is an area of of research and the work that you're doing that doesn't end. Right? But, you know, if all of, you know, this this research that, you know, we talked about this earth simulation on better understanding the climate. If all of this works, right, works and and continues, to improve with generative AI and everyone's models get better and we better understand extreme weather and climate, like, what does that mean? Right? Like, what does it actually mean for all of us? Is it just okay? Well, now we're safer and we just, you know, have a better idea of of the earth around us, but what does this ultimately mean, like, bigger picture as, you know, generative AI, you you know, really just changes, you know, what humans are capable of?
Mike Pitchard [00:30:17]:
Well, more reliable predictions, more accurate predictions are fundamental to planning, and we've got a lot of planning to do. You know, generative AI is not going to erase the fact that, you know, there's been 200 years of, of emissions that have changed the composition of the atmosphere for including a molecule that has a 100 year lifetime and therefore will will be remaining in the atmosphere for a long time. But but as the, as the weather changes, as the climate warms, I think that the ability to sidestep physics and to not have to rely on human assumptions about how, physics we cannot afford to represent numerically, but instead rely on data and AI to learn relationships. It it, you know, has the capacity to extrapolate beyond the boundaries of the observed record enough for us to get some benefits in extreme weather prediction that I think would be helpful to planning. You know, can I nerd out for one more moment, Jordan?
Jordan Wilson [00:31:08]:
Yes. Please. This is good. This is good.
Mike Pitchard [00:31:11]:
Okay. The the seeming act of video prediction in in in AI for weather where you're trying to predict the frame of ad, not of an image, not of 3 channels, RGB pixels, but rather of, you know, hundreds of channels, temperatures, winds at different altitudes, It it can appear like a an an imitation of video generation, but it's actually deeply physical. What's learned, what's trained along the way is a huge AI model that's learned physics. Now I was skeptical of this 2 years ago as a as a physicist physicist and a ML Libet compared to all my NVIDIA colleagues, but but I'm convinced now. And and the reason I'm convinced is because, a wonderful atmospheric science professor at the University of Washington took one of these trained AI models and then poked it and exposed it to to clean conditions that were never in the training data, the noisy data of real weather, but analogous to the kinds of pen and pencil problems we work on in grad school atmospheric science where you examine the force balances in response to idealized conditions. They proved the AI model could could produce beautiful clean solutions to these beautiful canonical problems. And it's very convincing evidence that there's learned physics. And and when there's learned physics, then there's the hope of generalization.
Mike Pitchard [00:32:18]:
And, yeah. So so I I really, I think we don't know the limits of that generalization, but it it could it could really benefit some of the existential problems we have in simulating the climate deterministically with physics.
Jordan Wilson [00:32:36]:
So much so much to wrap our heads around here, Micah. It's it's it's early in the morning for me. I'm a little bit tired, but this is waking my brain up. But, you know, as we wrap up here because we've covered a lot in a very short amount of time, maybe what is the one most important takeaway? And I know that's hard because we covered so much, but what is the one most important takeaway that you want, our our listeners and viewers, to understand when it comes to climate forecasting and weather prediction and how generative AI is changing all of that?
Mike Pitchard [00:33:10]:
Right. I I would say that, AI for weather prediction is here. The skill is unequivocal. Then the largest meteorological agencies are responding. The future of AI will include or the future of weather prediction will include AI, so stay tuned and watch what it appears on your on your app that you like to look at eventually. And the AI for climate prediction is coming, and that, multiple world world leading climate models are being infused with AI subcomponents. And, and and there's an enterprise of people working wonderfully in a collaborative spirit across the private and public domain in this sector, a problem that's really important for humanity. And, you know, there's some applications of AI that give people pause, but this is one I think we can all feel really good about.
Mike Pitchard [00:33:52]:
It's authentically quite important to Micah's future. And, it's a it's a a great privilege to get to work on it, with with such fantastic people at NVIDIA, and, really appreciate your interest.
Jordan Wilson [00:34:05]:
Alright. Well, hey. I I think this was an important one. A timely, you know, very timely for us to better understand with all the extreme weather we've seen recently, how generative AI and all the research that you're doing can really help us prepare, for, you know, hopefully safer, you know, safer day to day dealing, with the weather. So, Mike, thank you so much for joining the Everyday AI Show. We really appreciate your time.
Mike Pitchard [00:34:30]:
Thanks, Jordan.
Jordan Wilson [00:34:31]:
And, hey, as a reminder y'all, there was a lot there. Make sure if you haven't already, please go to your everydayai.com. Sign up for the free daily newsletter. Mike just dropped so many. He was just making it rain facts and figures on us. So we're gonna be recapping the most important takeaways in the newsletter. So if you haven't, please go to your everyday ai.com. Thank you for tuning in.
Jordan Wilson [00:34:52]:
Please join us next time and every day for more everyday AI. Thanks y'all.