Thank you. Hello, and welcome to the NVIDIA AI Podcast.
I'm your host, Noah Kravitz. One of the things that makes the weather so dangerous is that it's so hard to predict.
Tornadoes, hail, high winds, and flash floods cause billions of dollars in property damage, and injure or kill hundreds of people in the U.S. each year.
Knowing when storms may strike can save lives and property, and it's an area where deep learning is now being put to work.
Our guest is part of a team at the National Center for Atmospheric Research that's doing just that.
David John Gagne is a postdoctoral fellow at the National Center for Atmospheric Research. where he researches ways to use deep learning for severe weather forecasting and model parameterization.
David John, welcome, and thanks so much for joining the AI podcast.
Thank you. So you've been studying the weather for going on a decade now at least, is that right?
That is correct. You've worked as a student, a consultant, and now you're at NCAR.
So tell us about what you're studying, what you're working on, and how deep learning fits into that.
So for the past decade, I've been working on various research projects that are at the intersection of weather and machine learning.
Going back to starting with just using decision tree ensembles, things like random forests or gradient boosting regression to predict things like hail, where it's going to occur, how big it's going to be.
Also look at extreme rainfall, aircraft turbulence, and also predicting both wind and solar energy so that in a given situation when a Like a power company needs to know how much wind there's going to be over the next day or how much sunlight is going to be coming down so that they can balance their energy. wind and sun power with power from more traditional sources like natural gas or coal or Or are there less variable sources?
Right. So let's get into it. How accurate, or I don't know if accuracy is the best way to describe it, But how much confidence do you have right now in the models and the predictions that you're working with?
I have increasing confidence. There's still lots of room for improvement.
And the bar, with no disrespect to the local weatherman –
No, the bar isn't necessarily super high in the layperson's eyes.
Correct. But there are still – can be millions and billions of dollars riding on the line with some of these – predictions and how accurate they are.
So making them as accurate as possible is important.
We also are looking at a number of different baselines, including simplified models more physics-based models.
We're also even in some cases trying to compare with human performance and see if we are improving on humans and If we can work together with the machine learning model and the humans, can we make something even better?
So it sounds like you've got your hands, or you have in the past several years, in a bunch of different pots related... to the weather, you know, sort of broadly.
You talked about size of hail, you talked about energy coming from the sun and the wind, that sort of thing.
What is a typical, I don't know, day or perhaps project look like for you?
So for a typical project, I might be talking to say the national weather service or, or a private company that, that for instance, has some really specific need where weather is a major impact for them and they want to be able to take all the data they have and all the weather data that's out there, mesh them together and create some kind of end product that they can then use to make better decisions.
So that's part of my day. I'm often having to gather all the different data sources and bring them together. do a lot of pre-processing, basically trying to get everything in the right format, match up times between all the weather observations and the model output and maybe the private company's data, bring all that together, then Then it has to train all the different machine learning models.
And when we're doing this, we want to make sure that the machine learning models are actually learning what they're supposed to be learning.
In addition to the training part, we want to make sure that the models are evaluated correctly.
So we look at more traditional metrics. your root mean squared errors and areas under the rock curve.
But we also want to look at some of the physical realism of the model.
Is it picking up on features that are actual physical processes or is it picking up on some statistical anomaly in the data because someone just happened to running at a certain resolution or there's some other variable that's causing all these other variables to vary.
And we want to make sure all that is working in sync and giving us the results that we should be getting.
So I do a lot of visualization, all kinds of kind of analysis ranking, that sort of meta modeling of the actual machine learning model. to make sure that all this is working together properly.
And how far out in advance are you working or do you need to work?
If you had a client who came in with... I don't know, sort of an emergency building project they needed to get forecasting on for tomorrow or two days out or what kind of lead time do you usually work with?
In terms of the model running lead time, I've worked on projects ranging for everything from kind of now cast type situations where the model needs to be outputting like every hour or every few minutes to predicting a day or two ahead.
And I've even done a little bit of experimentation with seasonal to sub-seasonal forecasting so looking at one month to three months ahead and is there any skill there and it turns out there in some situations there actually is a little bit of skill if you bring in the right data Beyond that, in terms of actually developing these products, that's a months to years kind of timeframe.
Sure. There's obviously increasing numbers of off-the-shelf products, and I heavily use a lot of open-source tools, especially in the Python software stack.
And that really helps accelerate things and being and being able to Just take things off the shelf and bring them together and integrate it with all the weather tools is very powerful.
So let's go back in time a little bit and talk about your background for a second.
We mentioned at the top you've been Doing this for a while now, did you, as a kid or earlier in your education, were you fascinated by the weather?
Were you into computers and then data science and machine learning and the weather was just kind of a place to apply your skills?
How'd you get into this? So I really got into weather pretty heavily as a young child, basically when I was around kindergarten age.
There are a couple of major disasters in the US.
So there was like Hurricane Andrew was one of the first ones I remember being a huge impact.
There's also a big F4 tornado that hit near where I lived outside of Richmond, Virginia.
Those kind of crystallized weather in my mind.
And then over time, I basically read all the weather books and watched all the weather videos you could possibly...
I could possibly get my hands on. Then had a number of mentors along the way who kind of helped stir my interest further in whether And those people helped drive me toward going to school in meteorology.
And then, so I went to the University of Oklahoma for my undergrad and went into being my both master's and PhD.
When I got to Oklahoma- In meteorology? Yes.
Okay. Okay. Yeah, everything was in meteorology.
But by some happenstance, in my freshman year, I applied for a summer internship.
And I originally... When I applied for it, I didn't actually get into that particular internship program, but they – shared the applications from the internship program with a computer science professor who's also looking for basically someone to help work on her projects.
And so this computer science professor, Dr. Amy McGovern, she was interested in basically studying tornadoes with machine learning and looking at these giant simulations of tornadoes. and supercells and seeing how the patterns line up.
At this point, machine learning and AI was kind of in sort of a second AI winter in a lot of ways.
It was very marginalized at the time. Right.
What time frame? This is about 2007. Okay, sure.
Well, some of the... Building blocks of deep learning were starting to get going at that point.
And the broader consciousness that was pretty much in the background.
The weather area was... really not seen as a big thing.
So she recruited me to work on the summer project and then on doing storm classification.
And so I basically got up to speed on decision trees and clustering and really basic AI.
And then just kept working with her from there because she had funding to keep supporting me.
So so at that point, I learned this. up and coming programming language called Python start, started playing around more with, uh, bigger decision trees and ensembles of decision trees, and then started like sharing some of my work and really getting into it.
And I just kept going from there. And, uh, I had further opportunities to stay on as a grad student and expand to higher resolution modeling and, and, In that time period, I also started connecting with NCAR, where I had an opportunity to visit and then start connecting with the scientists there, which... eventually led to me being able to visit for a year and work alongside other meteorologists and other machine learning experts on on the solar energy problem.
And from there, I became I started as a postdoc and gotten and took that chance to to dive from decision tree work into deep learning.
Very cool. It's funny how those little twists of fates, like your original application wind up shaping things.
Our guest today is David John Gagne. He's a postdoctoral fellow at NCAR, the National Center for Atmospheric Research. where he's studying ways to use deep learning for weather forecasting, predicting severe weather, working with natural energy sources and all kinds of things related to the weather.
So along that journey, whether way back or even more recently, What, if anything, surprised you, either about the evolution of the tools you're using or the applications of deep learning, machine learning to studying the web? in particular?
There's been a number of, I think, big surprising things.
One, I think, is just recently how fast the interest in deep learning, machine learning has exploded just in the past couple of years.
For a long time, when I had first gotten started, It was more of a niche interest and I experienced a lot of skepticism from the more traditional meteorologists.
But recently with one publishing some positive results showing how the stuff can actually improve in a somewhat rigorous fashion on some existing prediction systems and running it and bringing some of these models in real time and seeing that the results do look physically realistic and do make sense for the most part i think that helped but also the broader consciousness of deep learning and basically just permeating the media uh has has led to a lot of people saying, hey, we should take a closer look at this now. big weather conference, the AMS, American Meteorological Society annual meeting,
Our AI conference, which we've been running since the mid 90s, usually had decent attendance, but not Not overwhelming attendance.
We were kind of off in the corner half the time, but this year our, our room was just overflowing.
We had, uh, for most of the talks, it wasn't just like a couple of talks.
It was like. Most of the conference, we had to turn people away in some cases because it was just crowded.
So that's a very encouraging sign. Absolutely.
Yeah. Yeah, and well-deserved after all the years of work.
You mentioned before Energy companies being a client that you work with.
What other types of organizations do you work with in your current role?
I've focused more on the government side, so I've worked a lot with the Storm Prediction Center, the National Severe Storms Lab on kind of my hail forecasting modeling.
And with that, just to mention, as you mentioned before, Are you working on or are you actually able to predict not just when hail would occur? uh perhaps you know how long it might occur but you actually talked about the size of the hail itself Yes, that's a key component to know both where hail will actually fall and then how big it will be.
So it's basically two machine learning models that can do that.
One that just predicts the probability of hail based off of where in the past the conditions are right for hail, but then it doesn't actually, it doesn't always happen. right.
So if you can match up where the conditions match it happen with, where the hail occurs.
That's one part of the problem. And then if you know that hail has occurred, you can build another model that will look at the distribution of hail sizes observed in area and predict that distribution.
And from there, you can look at that long tail and see if there's going to be an enhanced chance of really big hail.
What do you consider big hail? I'm kind of curious now because I've had some cars dented by decent-sized hail balls, but what's big hail?
So severe hail, basically the size that it's going to start denting your car, it's about the size of a quarter.
That's the size of a court, right? That makes sense, yeah.
Significant severe hail is usually about two inches, which is close to about a golf ball size.
Things get really, really bad when you're dealing with baseballs or bigger.
Yeah. What's and we're off on a tangent, but what's the largest size hail you've either encountered or dealt with in your work?
The largest I've seen personally, I think it's about baseball, didn't actually go out to to grab it because I was driving at the time, but then trying to get out of the way of it because it was a giant supercell above my head.
Where was that? uh this is outside of norman oklahoma uh so i i have storm chased uh in the past I didn't want to get off topic because you were talking about the nuts and bolts of deep learning here, but I've watched a few Storm Chaser videos.
I grew up for a while in Ohio, so I have memories of... you know, seeing tornadoes and hiding in the basement during tornado warnings.
But the videos I can remember seeing were kind of mostly in the, you know, some in Oklahoma and that part of the country.
So what's storm chasing like? It's very long periods of tedium and driving followed by a few minutes of sheer terror. would be one way to describe it.
Often you're basically having to drive hundreds of miles to go to the area where the storms are most likely to occur. with the exploding popularity of storm chasing, there are also a lot of other people out there.
So you're at this standing at this rural gas station in middle of nowhere, Oklahoma.
And then suddenly you see half your classmates also pull up at the same gas station.
They call it chaser convergence, but it's a good way to meet up with people and see friends.
Then once the storms do get going, usually you're having to drive and keep up with the storm, but also maintain a safe distance and have some situational awareness and make sure that that another storm isn't forming right behind you or that the storm that you're watching isn't going to suddenly cut you off.
And that can be surprisingly hard at times, especially if the road network isn't very friendly.
Right, right, right. I think it would be an interesting application of a self-driving car to see how well we're doing a storm chasing situation.
Oh, that is. Now we're on to something, I think, a self-driving storm chaser.
I like that. So you spoke a little bit about, we talked about hail and the variability of hail size and why that's an important factor in looking at hailstorms.
What are you working on right now that pertains to Hale in particular?
So right now I'm looking at how deep learning can learn the spatial patterns associated with supercell thunderstorms, which are the main producers of hail and trying to figure out if I can find spatial patterns with the deep learning model that relate to when big hail is going to be occurring.
So to do this, I basically build a simple convolutional neural network.
Simple. Simple for you. and compared it with some other more simpler machine learning approaches, just like a simple logistic regression and principal component analysis. and basically run them all on this large data set I have of 100,000 simulated thunderstorms and then the associated whether or not hail happened and how big it was and train all the different models ran a whole bunch of different parameter combinations and What I was able to find was that the deep learning is significantly improving on the accuracy of the hail forecast.
By looking at the individual layers of the deep learning model and visualizing what inputs activate those layers I'm able to see that the neural networks are actually learning physical features so they're learning things that meteorologists would look for, like seeing wind turning with height or having air flowing inward at the lowest level and then going up in the storm and seeing conditions where basically hail seeds or grapple from one storm actually being invected into another storm.
The network learned how to find that without me specifying anything in that, in that regard.
That is wow. Okay. So tracking, I mean, it made me think of, uh, you know, uh, a bee pollinating flowers and tracking. the hail seeds is that the term you use seeds embryo would be the more accurate term but a seed might be another way to think of it i like the living aspect of embryo yeah Yeah, the embryos basically can grow in what's called the updraft of the core of one storm, but they don't get very big there or they get thrown out. if they get too close to the center of the updraft.
And then they get shot through what's called the anvil of the storm, basically the high clouds that go ahead of the thunderstorm.
And if there's another storm, ahead of that those those embryos can fall into the after after the other storm and then Because they're already somewhat big, the process has been kick-started, so the other storm can use that to grow it even bigger and have a much larger set of hail.
And then there's a lot of heating and stuff that goes on that can help strengthen that storm further.
So all kinds of fun nonlinear processes.
Right. And your networks have learned to pick up on some of that and track it.
Yes. That is fantastic. And to figure out that the network is doing this, I've used some of the techniques similar to Deep Dream and some of those psychedelic dog... and other picture transformation images.
But it turns out it's also really good for interrogating networks for its scientific content.
Right, right. Excellent. So kind of looking sort of more back to the present and kind of looking forward How do you see, and not to put you on the spot to brag about your own work, but the types of work that you're doing, your colleagues are doing?
How do you see this kind of going forward to the next five or so years, let's say? of predicting severe weather, being able to help, whether it's a governmental agency or whoever it may be to prepare for big weather events and, you know, not to get into talking about the weather itself too much, but certainly in... let's say the mainstream media over the past couple of years, there's been more attention on kind of significant weather events that have happened.
Where do you see this all headed? Even what are you hoping to do with your work over the next few years?
So there's probably a couple of big areas where I'm really interested in working further.
And continuing the development of decision support systems for meteorologists or the general public that kind of take in all of the numerical model output and observations and remote sensing data from radars and satellites and merge that together to make the best estimates of where hail or tornadoes or any of these other severe threats are going to occur.
And there's a lot of... And there's a huge amount of data that's available now.
And a lot of it is just sitting on hard drives not being looked at.
So... If we can train our machine learning models or deep learning models to be able to find these patterns for us, then that data will be put to good use.
And if you can get all that running in real time, then the forecasters will have a better framework where they can immediately look to where the threats are and analyze it and see and have all that information available so that they can make the best forecast or But even going further beyond that, someone with their cell phone needs to know there's going to be hail or tornadoes nearby or there's an enhanced chance of that.
If you can bring all that information to bear and push that all the way out to them.
The other area where I'm really interested is actually going back not just looking at the end of the weather data pipeline, but getting within it and some of the key components.
So one of the things that are kind of core to our current weather forecasting ability are these numerical weather models are basically solving the equations of motion and moisture and thermodynamics at every grid point all over the Earth.
And these are giant models. They're They run on massive supercomputers every day, multiple times a day and run out to many days in the future.
And they're really expensive, but there's a lot of simplifications that had to be made in them so that they can run in a reasonable amount of time.
All of those use fairly simple models, but I think there's room for deep learning and other kind of advanced machine learning techniques to replace some of those. process models and inside the bigger model and basically either help one emulate the a more complex process model but just run a lot faster, or even take all the data that was used to generate the simpler models and use it to build a more complex model I can more realistically simulate the conditions that are going on and give us better estimates of what's happening at the surface, what's happening in clouds. looking at basically turbulence and all kinds of other factors.
And then if you get those processes right, then the model output will be more accurate.
So you won't have to do as much post-processing to get good results.
And then you can also run the numerical model.
You can run more versions of it. run larger we call ensembles so so we can get a better estimate of the uncertainty and what's going on right I don't know how much of this you look at in your work, but I'll throw this out there anyway.
Has the weather changed a lot over time, over the past 20 years, 40 years, 100 years, 500 years?
Yes. The average weather, the climate has in fact changing and based off of all of the scientific evidence has been gathered at places like NCAR and at other research institutions around the world.
We can clearly tell that the climate is changing and humans are playing a pretty big role in it.
And how the climate is changing has impacts all the way down from... say everything from agriculture and like if we're gonna have droughts or more floods, that's a, that's a big problem.
For the signal related to extreme weather, it's not as clear in some circumstances, especially like saying there's more hurricanes or fewer hurricanes.
There's some evidence that... In the future, we may have fewer hurricanes, but the ones that do happen will be stronger because there'll be more moisture in the air because it's warmer.
And so that may help them basically Reduce more rain, have stronger winds, cause more problems there.
There's also some evidence that at the thunderstorm level that there may be more variability so some years we may not remember already starting to see this a little bit in the record but some years you barely see any tornadoes but then other years you have giant outbreaks and you we might have more year-to-year variabilities because of how the climate system is changing.
So It's a lot to think about. A lot for the GPUs to process, that's for sure.
Definitely. Well, David John Gagne, thank you so much for joining us again.
Postdoctoral fellow at NCAR, the National Center for Atmospheric Research.
If people, if listeners want to find out more about the work. you're doing and NCAR is doing, anything else related to storm prediction and using deep learning?
Is there somewhere online they can go? I highly recommend checking out the UCAR website, ucar.edu.
We have a lot of resources there. I'm on Twitter at DJGarnierDose.
Feel free to contact me from there. Excellent.
Well, thank you so much. And at the very least, I hope there's sunny weather for you the rest of today.
Thank you.
