How do we predict the weather?

24:06So you can stay vital, stay you. Visit VitalProteins.com to learn more and where to buy. These statements have not been evaluated by the Food and Drug Administration. This product is not intended to diagnose, treat, cure, or prevent any disease.

24:29All right, and we are back. So now we're going to talk about the kinds of data that we use to make weather predictions. and I'm going to bet it involves satellites. Always going with space first, right? Yep, yep. It does involve satellites, but there's an amazing, incredible variety of data sources we have to understand the weather. And yet still, it's not nearly enough, right? As you'll hear, our weather prediction would be so much better if we had more data. We're really limited by the data. But we have lots of different kinds. We have weather stations on the surface. And so a lot of these are called like automatic weather stations that are scattered across the country.

25:07They're just basically a bunch of sensors with a battery and like either a wind turbine or a solar panel to get power. And they measure things like temperature and pressure and wind speed and, you know, precipitation. Just the raw measurements you need to know, like what's going on out there. What is the state of the weather right now? Because again, if you want to predict the future weather, you've got to know what's going on right now. So is this like a citizen science thing where like I could purchase one of these weather stations and hook it in to what's happening at like the national level?

25:39Yes and no. So there are a few sort of official stations. There's a bunch of different networks. The highest quality ones, there's like 10 ,000 of these scattered around the earth and they're operated by weather services and government agencies. But there's a bigger network of like quarter million of these things. Some of these are personal weather stations that, yeah, people just build and publish the data. And there's an amazing network. It's called COCORAHS, C-O-C-O-R-A-H-S, Community Collaborative Rain, Hail, and Snow. Wow, okay. And you can just build your own device and add it to the network and contribute.

26:15And I think that's super awesome because it's definitely limited by the data we have. One problem is that these things tend to be where the people are. We have a few top of Mount Washington or whatever, but mostly these things are put up by people where people are near. And so there's lots in India, for example, but very few across Siberia. And often the best ones are places like airports. Airports really need to know weather, so they have excellent weather stations. But the weather at LaGuardia is not the same as the weather in Manhattan. And so often the airport weather stations are very, very precise and used heavily in the models, but they're not giving you the measurements exactly where you want them to be.

26:56Okay. So is that a problem for just the people who are in areas where there's not enough weather detectors? Or is that a problem for all of us? Because what's happening in Siberia is important to what's happening in India. Yeah. What happens in Siberia doesn't stay in Siberia, unfortunately. It contributes to uncertainty and error across the model. and the earth is one big system, which is why you can't just be like, I'm only going to predict the weather in Manhattan. I only need to think about Manhattan. You need to model the whole planet in order to get the weather in Manhattan. So yeah, absolutely.

27:29And that's why we have lots of different kinds of sensors, not just these automatic weather stations. We also have things like weather radar, and you might've seen these on your local weather channel, like let's look at the Doppler. This measure of precipitation, it also measures the velocity of those raindrops. And this is a really cool story because it comes out of World War II. It's another example of reusing military technology and infrastructure after World War II to do some science. Thank you, war. Ooh, boy. Hot take. Pulling it back. Well, there you are again, finding the silver lining.

28:04Tens of millions of people died, but we have better weather predictions. So the way radar works is that it sends these pulses of microwave radiation. The wavelengths are like one to 10 centimeters. That's the microwave region. And it sends a pulse for like a microsecond. And then it listens for return signals. So like it sends this pulse and rain drops will reflect. So it gets a signal back and it listens for like a few milliseconds and then sends another pulse. And so it can tell where the clouds are and it can tell the velocity of those clouds by the change in frequency. This is the Doppler effect, right?

28:41And this is exactly the same effect as like stars are moving away from you. So their light is red shifted. When the radar pulse comes back, if the frequency is shifted, you can tell which direction that raindrop is moving. So that sounds complicated because like there's not just one raindrop out there. There's a bunch. And so I can imagine like your pulse getting lost as it bounces off of multiple raindrops and doesn't make it back to you. What am I missing? This sounds hard. No, it is hard. But you're not detecting individual raindrops. So you're detecting clouds mostly, like which direction is this cloud going?

29:14And, you know, initially this was a problem because in World War II, radar operators were trying to use radar to discover like enemy planes. And they noticed like, man, clouds are getting in the way. And then other folks were like, oh, wait, you can use radar to see clouds? Awesome. And so then after World War II, they started using this to measure the velocity of clouds and to see them. And there's this moment in like 1961 when Hurricane Carla was approaching the coast of Texas and Dan Rather went down there to a weather station and they were using radar to see the clouds and to see their direction.

29:47and he had them draw like the coast of Texas over this image of the hurricane that showed everybody like, wow, this is a massive hurricane moving fast towards the shore and probably saved thousands of lives because he publicized this like incoming storm much more rapidly than we could otherwise without this kind of technology. Wow. So this, yeah, this weather radar is really helpful. Do you think it still has the same effect or do you think people are just kind of like, oh, there's hurricanes, I've seen them before, they get big and they don't always leave? People don't always leave. There's always somebody who's going to ride out the storm, right?

30:20Yeah. And I don't know about the psychology there, but at least now we can inform people further in advance and let them know where these things are likely to go. But there's still always uncertainty. And we'll talk about that in a minute. You don't just have one weather prediction. You have an ensemble. You have an envelope of predictions because you don't have perfect data and you don't have a perfect model. And so often what you do is you vary your data a little bit within the uncertainties and run the model again. And then you get a different prediction. And it'll give you a sense of the spread of the possible outcomes.

30:49So you might see when there's like a hurricane approaching the coast of Florida, they have a bunch of possible trajectories. Those are all like different runs of the weather model, assuming different initial conditions because we have uncertainty. We don't have perfect data. I personally really enjoy learning about the uncertainty like in life in general. And whenever I look at those, I have this weird feeling of like security. Like they figured it out and they know what the errors are. We're good. We know what to avoid. So maybe that's a little bit giving it a little too much credit, but it's still amazing.

31:21And another really important source of uncertainty in our models is what's happening in the ocean. Like how hot is it? How cold is it? How are things circulating? All this kind of stuff. And so we need data about the ocean, but not a lot of people live in the ocean. So we don't have like these automatic weather stations, but we do have buoys. These are like floating weather stations. and around the world there's a couple of thousand of these depending on the type that have these like temperature sensors on the surface but we also have this hilarious data from what's going on deeper in the ocean that historically has come from people on ships taking a bucket dropping it into the ocean pulling it up and then measuring the temperature of the water and it's like really that lo-fi but for many years that's all we had we had like no other way reliably to know how cold is it in the ocean.

32:14And this is an example of like, it's not just data. You need to take data and interpret it and clean it and correct it. And I spoke to an expert here at UCI, Jane Baldwin, who told me that like, you had to correct for like how long the bucket was out of the water before they dunked the thermometer in it and how Japanese ships and US ships use a different bucket and it had different effects. And like, you got to really know, you got to be an expert in how this data was taken and what it really means. Yeah. So for a while I was doing some water quality work and we had this like tube and you would put the tube underwater and then you'd sort of press a button and like caps would pop into place on both sides of the tube.

32:51And then you could lift it up. And so you could get a water sample from specifically different depths. And it was always kind of fun to use that device. Yeah. And you might think like, that's ridiculous. What a silly system. And it's a little bit silly, but if it's the only data you have, it's better than no data. Yeah. Right. As long as you understand the uncertainties in it. And my friend Jane was telling me that misunderstanding this data might be a cause for some weird pauses in global warming trends, that it could just be like a misinterpretation of this ship bucket dunk data. Oh no. I know.

33:24We're so lo-fi. It's amazing. These days, we have these cool robotic floats that float on the surface of the ocean and then dive down up to 2 ,000 meters, measure things down in the ocean, and then come back up and beam it to satellites or whatever. So we're getting better, obviously. But what's really valuable is longitudinal data. You want data as far back as you can so you can understand bigger trends. So you can't just say, oh, that ship bucket dunk data is ridiculous. Let's ignore it. It's the only data you have for like 30 years. And so trends in that data do tell you something. Very cool.

33:59Okay, so now we've gone down deep. How do we get data from up high? Yeah, because the weather's not just at the surface, right? And the weather folks call the surface the two meter level because they want to measure the temperature not on the ground, literally, but like two meters up, like where your head is, essentially. But they also need to know what's going on even further. So we take measurements in the upper atmosphere. We use weather balloons. And these are literally what you imagine. And you put like a bunch of helium in a balloon and you put a weather station on it that can measure altitude, pressure, temperature, humidity, wind speed, etc.

34:33And you just let it go. And it rises because helium rises. And as it goes up, the balloon expands because the pressure in the upper atmosphere is less. And eventually it pops and then the thing comes back down. So these are like one-time uses. And they can go up like 20 kilometers. When I visited St. Catherine's University in Minnesota to give a talk, they had a special day where they launched a weather balloon for my visit and did a demonstration for all the students. And it was the coolest thing ever. It was awesome. Super cool, right? Yeah. These are amazing experiments. And I know people who do physics experiments on balloons where they go to the Antarctic and they let up a balloon and it floats in the atmosphere for up to a month or something.

35:16Wow. That's really brave work because you spent like four years building this instrument and then you're putting it on a balloon and sending it to the atmosphere. And sometimes it's just gone. Like it just disappears and you lose your whole thesis. And this seems like kind of bespoke, right? And it is. There's like a couple hundred launches per day in the United States, but it's not reliable. It's not like the place you visited, they do exactly the same balloon launch every single day at the same time, right? Which is the most useful thing for a weather model is like reliable data. and we don't have a lot of them.

35:47But again, this helps you probe the upper atmosphere. We don't have many ways to measure the temperature in the upper atmosphere. This is a really powerful one. Do we also use planes? We do use planes because every airplane you've been on has really valuable information about weather because it samples from the two meter level up to like 30 ,000 feet. And aircraft, of course, have sensors to measure wind speed and temperature and all this kind of stuff. So every commercial airplane has these sensors, collects this data, and then sells it to the government. NOAA buys this data because there's so many flights.

36:22Look at a map of airplane flights for a single day in the United States. There are so many flights. They crisscross the country, and it's incredibly valuable data. This is usually very high-quality data because it's very important for these planes to understand the weather. I had no idea NOAA was getting access to all of that data. That's super cool. Yeah, it's super cool. Basically, any way you can imagine to learn the state of the weather somewhere on Earth, somebody's doing it. Because the more data we have, the better these models get. But then, of course, we can go all the way up to space, right?

36:54Because there are places where there are no automatic weather stations and there are no buoys and there are no airplane flights. Yet they still contribute to the weather prediction in Kansas or in Mexico City or whatever. And so we have satellites. And since about 1979, we've had weather satellites. We, of course, had satellites earlier than that, but none devoted to gathering weather data. And these primarily cover things like storm systems and cloud patterns. They can tell you where the snow is. They can also tell you where wildfires are, which is an important part of the weather. Oh, yeah. And so they can't directly measure what is the temperature in Houston right now.

37:31But they can make indirect measurements. Like, for example, they can measure the amount of infrared radiation from the surface, and that is connected to the temperature, but it's actually connected to the temperature of the surface, not the two-meter level. So, like, how hot is the blacktop in Houston right now? That's what your satellite is telling you. And you have to use that to infer how hot is it two meters above the blacktop in Houston, which is what you actually want to know. That sounds hard. It's hard. Yeah, exactly. And so we also don't have a lot of satellites because they're expensive.

38:04There's something like 20 satellites, a mixture between geostationary and polar orbits. Eight of them are operated by NOAA, but there's a bunch out there. But my friend the climate scientist says that we might be on the verge of having a lot more data because launching stuff into space is cheaper. And now we can do like small satellites, CubeSats. These might give us more data, but not as high quality as like the dedicated, super nerd-designed billion-dollar satellites. On the other hand, we don't know what the future holds for supporting and operating these satellites. This requires money to fund these things and have people interpreting these things.

38:41We don't know how long the government is going to continue to support it. They could just unfund this stuff or turn off weather stations. Oh, my gosh. It's more than just like, oh, we turned it off for a year. having continuous records is super important for these models for predicting the immediate weather, but also for the long-term climate models, which are essentially an average of the weather. And so even turning it off briefly could be very damaging for our ability to do long-term predictions. And I'm kind of blown away by the fact that we only have 20 satellites. I guess I had assumed since, you know, there's like 5 ,000 satellites up there or something, I guess most of them are dedicated to like beaming cat videos to us from anywhere in the world.

39:21But like, weather seems so important, you know, for farmers, for like people who are traveling, just like for everything. Yeah, that's true. But the satellites don't give you a direct measurement of what you're most interested in. They're essentially like really good for filling in the gaps for places where you have no other measurements. So yeah, it would be great, but they're also super expensive. So you'll hear at the end, I asked one of the climate scientists, I spoke to, like, if you had a billion dollars, what would you spend it on? And satellites is not their top priority. Huh. Okay. All right.

39:50So maybe 20 is the right number. So you have all these different kinds of data. You have automatic weather stations. You have radar. You have buoys. You have ship bucket data. You have weather balloons, aircraft. You have satellites. What you need for your model are the initial conditions. What you need for your model is a set of what is the temperature and the pressure and the humidity everywhere on the planet right now so that I can run it and predict it in the future. And there's not a trivial step from like, here I have all this data to what are the initial conditions? Because the data can disagree, right?

40:22You have multiple measurements sometimes nearby using different kinds of sensors. How do you incorporate that? How do you clean this data? How do you decide what to use? How do you merge all of this into your best prediction? And so there's a lot of work in this area. It's called data assimilation of running sort of mini models, fluid dynamics to do like physics informed interpolations between the places where you don't have measurements and to factor in the various uncertainties from the various different kinds of measurements. So sometimes you like back up the model a little bit and feed in some data and then use it to predict the current initial conditions before you go to your full model.

41:01And then you do what we talked about earlier, which is ensembling. You say, well, here's my best guess for the weather like right now before we even run the model, but I'm going to make like 100 versions of it. And each one, I'm going to tweak my assumptions a little bit so I get an envelope where I hope reality somehow is described by one of these models or is near one of these models or the spread in these models describes my uncertainty in the state of the initial conditions. We haven't even done any predictions yet. This is just like measuring what's happening now, right? And what's so stressful for me to thinking about this is Like your data are coming in constantly.

41:36And so it's not like you do this once and then you're like, okay, good. Now we will project. It's like every second more data are coming in. So this has to be like a constant process that's happening over and over again and integrating the information into bigger models. And it's amazing. Exactly. And yeah, we haven't even talked about how the models work. And so let's take a break. And when we get back, we'll talk about how those models work.

42:06This July 4th, come celebrate at America's Block Party, hosted by America 250. America's Block Party is a can't-miss 4th of July concert happening at the Los Angeles Memorial Coliseum. Experience music performances by major artists, patriotic tributes, and the kickoff to Giving Fourth, helping to make July 4th the largest day of giving in American history. It's more than just fireworks. Join this landmark celebration and get your America's Block Party tickets now for$17.76 at america250.org slash LA. Introducing the all-new Mazda CX-5. Featuring more connection. Hey, Google, where's the nearest Pilates class?

42:44Safety that has your back. More discovery on the scenic routes. More passion in the details. And more control in changing weather. The all-new Mazda CX-5. More to move every side of you. See it in five films at mazdausa.com slash five sides. Google is a trademark of Google LLC. Sequences shortened and simulated. Eczema is unpredictable. But you can flare less with EpGliss. A once-monthly treatment for moderate to severe eczema. After an initial four-month or longer dosing phase, about four in ten people taking EpGliss achieved itch relief and clear or almost clear skin at 16 weeks. And most of those people maintain skin that's still more clear at one year with monthly dosing.

43:26EbGliss, Librikizumab, LBKZ, a 250 milligram per two milliliter injection is a prescription medicine used to treat adults and children 12 years of age and older who weigh at least 88 pounds or 40 kilograms with moderate to severe eczema. Also called atopic dermatitis that is not well controlled with prescription therapies used on the skin or topicals or who cannot use topical therapies. EbGliss can be used with or without topical corticosteroids. Don't use if you're allergic to EbGliss. Allergic reactions can occur that can be severe. Eye problems can occur. Tell your doctor if you have new or worsening eye problems.

43:54You should not receive a live vaccine when treated with EPCLIS. Before starting EPCLIS, tell your doctor if you have a parasitic infection. Ask your doctor about EPCLIS and visit ebclus.lily.com or call 1-800-LILY-RX or 1-800-545-5979. Aging is real. And so are the benefits of adding Vital Proteins Collagen Peptides to your daily routine. New Vital Proteins Collagen Sparkling Water. Your daily glow-up now in three fresh flavors. Strawberry Blossom, Lemon Lime, and Blood Orange. Improved skin health in as little as 30 days thanks to collagen peptides. Cheers to that. Or go with our classic collagen peptides.

44:27So you can stay vital, stay you. Visit vitalproteins.com to learn more and where to buy. These statements have not been evaluated by the Food and Drug Administration. This product is not intended to diagnose, treat, cure, or prevent any disease.

44:47All right, so now we have all of this data, and you've got it into an ensemble, and you sort of maybe know what's happening right now, plus some uncertainty. How do you now predict what's going to happen next? Yeah, so it's simple. You just break out your pencil and paper and you do a bunch of string theory calculations and that's it, right? It's just like physics it into the future. Finally, string theory is useful. Yeah, unfortunately not. As we said earlier, like you can't describe everything. You have to make assumptions about what you're going to calculate and what you're going to simplify.

45:20Otherwise, you're never going to be able to make a prediction. right? Or your predictions are going to be done in a thousand years for the weather that's happening in an hour. And that's not useful. And so it's always a question of how to judiciously make those assumptions. So the current state of the art for weather modeling has basically two big pieces. One is directly model the atmosphere itself as if it's a big fluid. So use like Navier-Stokes equations and think about how it flows and how temperature moves through it. That's the like dynamical core of the model. But there's a bunch of stuff that influences the atmosphere that you don't explicitly include in the model.

45:57The clouds, the convection, the ocean, the radiation, the surface temperature, all this kind of stuff. Your model doesn't explicitly include that stuff. We don't have like a complete model of the ocean or the clouds, et cetera. And so we have like various inputs to this core piece that they call parametrizations that like capture the big picture effects of all these pieces that are not included directly in our model, but are influencing us. So I feel like this is a question where you think to yourself, am I about to ask a really stupid question? But I'm gonna move forward because that's my job in this podcast.

46:33The atmosphere is not a fluid though, right? And so like, am I wrong? So why are we modeling it as a fluid because we just can't model it as something else because it's too complicated and fluids are a simplification? Or Daniel, I don't think the atmosphere is a fluid. Well, it depends on what you mean by a fluid. And when it comes to how things flow and pressure, et cetera, the fluid dynamic equations do describe the atmosphere. And so fluid doesn't mean liquid, right? Fluid is about how things flow and move, right? So for example, the mantle of the Earth is a fluid. It flows. It's not a liquid.

47:19It's this weird solid-y kind of state and it moves, but it also flows. And so you can describe it, and it has convection. You can describe it with fluid equations. And so the Navier-Stokes equations are these famous equations that describe fluid dynamics. And they're pretty good at modeling the atmosphere. They're not perfect, right? They're not perfect, but they're pretty good at it. So yeah, I think fluid is not a liquid. It's just things that flow. Okay. In my head, fluid is synonymous with liquid. And so I have learned something today that will probably help me not look silly in the future.

47:50That's great. No, it was a great question. And so this is the big picture. You have the dynamical core, and then you have these parametrizations, and we'll dig into that. And we'll describe sort of the US approach to it, but there are three sort of major weather communities. There's the US, the UK, and the Japanese. and they have slightly different approaches, which is good because, you know, different predictions can cross-check each other. But then some people think it's bad because, hey, let's pool all of our resources and make one big global model. And that's awesome. But then you only have the one and you're not sure maybe it's all wrong.

48:22There's a lot of debate about, you know, how to deal with global questions and global resources. But who's the best? Oh, I'll give you a ranking at the end. Okay. All right. So the dynamical core, right? Think of the atmosphere. we're going to treat the atmosphere basically like a spherical cow, right? It is a spherical fluid, right? The atmosphere. I guess this is physics. Yeah. It's a thin shell around the earth and you know the temperature and the pressure, and then you can describe how it's going to flow, how the temperature and pressure are going to change using the Navier-Stokes equation.

48:56So Navier-Stokes is a set of really gnarly equations. They're nonlinear partial differential equations. A differential equation is one where like the value depends on how quickly it's changing. For example, like in ecology, you have differential equations that describe like predator and prey, right? These two things are coupled. And so these are nonlinear partial differential equations, which means like they depends on things squared or cubed. All that to say, they're very, very difficult to solve. In fact, differential equations in general are hard to solve. If you've taken a differential equations class, it's basically like differential equations are not solvable except for these four examples that we have answers to.

49:36And we know how to solve them. And so you just got to memorize those. It's a little bit like chemistry, I got to say. Oh, no. It's mostly unsolved, right? And the Navier-Stokes equations, we've known about them for like 200 years. They were initially developed to try to answer these questions about like how do things flow and how does momentum and mass flow through pipes, et cetera. Essentially, people took Newton's second law, F equals MA, and applied it to fluids and then added terms for stress and pressure and viscosity. And it's like a real triumph that we can describe this at all. But calculationally, it's a real bear.

50:12You can't sit down and derive a solution and say, here's my pressure and temperature. Let me crunch it through the Navier-Stokes equation. It's going to give me a formula. It's all numerical approximations, which means it takes a lot of computing to go from now to one second from now or two seconds from now. And that computing means approximating things. You're like doing numerical derivatives instead of exact analytical derivatives. Okay, so some of that got pretty complicated. But I guess what I want to know is when this is all done, I feel like if we are trying to model fluids, does this just tell us that like the wind is now over here going this fast?

50:49but before it was over there? And how many are we going to get to like how you get from that to like, and it's raining? Because that seems like a different problem, sort of, than how fluid is moving around. Yeah. So there's a couple of things to know there. You're exactly right. It takes the current conditions and tries to predict the future conditions. And those conditions are pressure and temperature, wind speed, humidity, right? These kinds of things. But because we're solving this numerically, we can't solve it everywhere. If you have a formula, an analytic description you can write down for like, where is my ball as I've thrown it?

51:21I can write that down on a piece of paper. I have a formula that can tell you where the ball is at any point in time. You ask me any point, literally any value of T, I could plug it into my formula and give you an answer. But if I don't have a formula that's called an analytics description, if all I have is a numerical estimate, then I've made a grid. I've said, I'm going to sample it at time one, time two, time three, time four. I'm going to make an estimate at those times. And I don't have an answer everywhere. And that's the situation we have with weather is that they put a grid on the planet and they estimate what's going to be the weather temperature, et cetera, in a grid of points, not everywhere over the planet.

51:58And you might think, oh, I bet that grid's pretty small, right? Maybe they measure it down to the centimeter or something. No, the grid sizes are like 10 kilometer cubes. What? Yes. That's too big. It's too big, right? They are averaging the temperature and the humidity over cubes of atmosphere, 10 kilometers on a side. It's crazy. And you might think that's way too big. On the other hand, that's still a lot of cubes, right? Like the atmosphere is a lot of 10 kilometer size cubes. And then the time steps are tens of minutes. Okay. Right. And this is awesome that we can even do this. It requires massive supercomputers we'll talk about in a minute.

52:36But the problem is that it ignores a lot of little details, like how big is a cloud? Usually they're like a kilometer or less. And so you're missing out on a lot of stuff by making your grid. Anything that happens that's subgrid, that's crucial and important, but is smaller than the size of your grid is not being described by your model. But your question was like, is this directly outputting like, hey, it's going to rain on Kelly's picnic? In a sense, yes. The direct outputs are things like temperature, pressure, humidity. And those are enough to tell you like, okay, it's going to rain because the pressure and the humidity are above some threshold or whatever.

53:12So it's not directly outputting like three centimeters of snow. There's another step you have to take after that, but it feeds into that. So those are the inputs you need to the next step, which says how much snow is going to fall. Okay. So let me see if I can do a super simplified version of this. You get all of the data that you have about a square in the grid and you do the best job you can to sort of summarize it and ensemble it and then you put it in the model. The model runs through the equations. Then does the information from the surrounding grids feed into your grid as well? Because you would expect there to be similarities between closely related squares in the grid.

53:51Yeah, you can't solve one grid at a time. You have to solve all the grids. Holy cow. Because the grids touch each other and influence each other and wind flows, right? Right. And that's why Siberia affects Manhattan over time because you've propagated these things from grid cell to grid cell. Absolutely. So does Siberia have bigger grid cells or just the same number of small grid cells, each with poorer data in them? Yeah, great question. So some of these models are adaptive, right? They have bigger grid cells where you have more uncertainty and smaller where you have more data. The most precise ones are the UK supercomputers.

54:25They go down to two kilometers in some cases. And some of them are like fixed grids and some of them are adaptive. Exactly. It depends a little bit on the model. But there's lots of details that are not described here. And these are called the parametrizations, like especially subgrid stuff and exchanges with other parts of the system that are not just the fluid. And one important thing are the clouds. You cannot model every individual cloud because clouds are smaller than your grid size. We do not have the compute to do that. People have tried. And you can do dedicated runs on subsets to try to resolve clouds.

54:59but then you don't have enough computing to do like ensemble. So you'd be like one prediction. You're like, well, here's a prediction, but I don't know what the uncertainties are on it at all. And so instead, what you tend to do is parametrize the bulk outcomes, the vapor, the clouds, the liquid, the ice, the rain, the snow, et cetera, the condensation, all this kind of stuff. You try to grab all of that average over what's happening in that grid cell and use that to inform your Navier-Stokes equation. So things you're not explicitly modeling, you're sort of like averaging over. You're losing all the details and saying like, well, on average, this is going to be the effect of clouds on my grid cell.

55:37And do you think that as, oh, well, I was going to say, do you think that as we continue to have more and more computing power and more and more supercomputers at some point, we'll be doing better here? But we were just talking about how Moore's law, we've maybe hit the end of that. So are we like, is this as good as we're ever going to get at it? This is probably an end of the podcast question, but I'm thinking it right now. No, I think that there's lots of possibilities for making this faster and more efficient and not just wait till computers get faster. Okay. There's definitely clever ideas and we'll get there.

56:09Yeah. All right. But there are lots of parts of the weather that are not directly described in the dynamical core. So not just the clouds, but also things like convection, like vertical transport of heat, you know, So especially this is complex boundary mixing near the surface, like the lowest kilometer or so of the atmosphere where you have like heat from the surface and turbulent momentum exchanges as wind is like hitting mountains and stuff. These things, you can't model all of those details. And so you have like parameterization schemes that model the turbulence and the boundary level mixings.

56:43There's radiation from the surface also, right? That changes from day to night. You have models of vegetation and snow, how those things couple. But then the biggest one is the ocean. We would love for our models to include also a Navier-Stokes simulation of the whole ocean. Might as well do that because the ocean plays a big role. But we don't have the compute for that at all. So we just use a slab ocean model. We just say, let's just assume the ocean is simple and we have a certain temperature and we assume how the energy transfers from the boundaries. And it's really quite simplified. But we're just limited, right?

57:24We don't have great data in the ocean and we don't have the compute to also model the ocean as well as the atmosphere. So places where we don't do our best approximations, which is like Navier-Stokes equation of the atmosphere, we have simplified versions, which are called parametrizations. they feed in to the core. But in the end, you got to take it to the computers. And this is why we have like massive supercomputers to make weather predictions. So NOAA in the US has a couple of really big facilities. They're called Dogwood and Cactus. One of them is in Virginia. You're welcome, everyone. And one of them is in Arizona.

57:57And they're huge, amazing computers. Those two have like 12.1 petaflops. You made that word up. Sounds like a made up word. Peta means quadrillion and flops are floating point operations. So, you know, it takes the computer time to add like 3.91 to 14.42 and floating point numbers, those numbers with a dot in them, right? Not integers are more computationally expensive to add or subtract. And that's what most of these models do, they're like, add this number, multiply by this. And so this is like the way you measure the speed of a computer. And so these computers can each do 12.1 quadrillion floating point operations per second.

58:41Imagine the guys back in the 1920s, they're like adding two numbers. It probably takes them a minute, right? Or if they're super good, it takes them 20 seconds. The computer does 12 quadrillion of these per second, right? So together with all of their computers. NOAA has about 50 petaflops, and that's what it uses to run its model. And so that's the state of the art in the United States. The Europeans have a couple of computers. There's one really big one in Bologna called Bullsequania, and it has about 30 petaflops. But the biggest, most powerful weather computer in the world is in the UK. It's at the Met Office, and it's built by Microsoft and it has 60 petaflops.

59:24And this is why the UK has some of the best weather prediction in the world. Because they have the biggest computer. They beat us. Yeah, exactly. They just spent more money. They bought more compute. This is literally like money equals computing power. I was tea drinking bastards. A day. Yeah. Well, you know, they got tricky weather over there and so they need it. It's an island after all. Good job guys and gals. The Japanese have a big investment in weather prediction computers also. There's one called Prime HPC. It has 31 petaflops. So these are really powerful devices and they run these huge models.

1:00:00And think about what the model does. It predicts the state of the atmosphere on these pretty chunky grids, but it's still, it's a huge amount of data. Like every few minutes, every 10 kilometers. My friend Jane was telling me that sometimes the data is so big that you just throw it away. You run it, You get like a summary number, but you can't keep all of the data because it would just like fill up all of the hard drives everywhere. And this is familiar for me because like at the LHC, we also we run these experiments every 25 nanoseconds. We throw away most of the data from that because it would just fill up all of our storage.

1:00:35And they're in a similar situation. They produce more data than they can store. So are these facilities where the Navier-Stokes equations are being run, or are these facilities where you have the output from each grid and now you are translating that into information about where the rain is falling? Both, yeah. Okay. So these programs do the data simulation, come up with the current initial conditions, and then also run the model forward to make those predictions. And from that, glean things like weather details, snowfall, et cetera. And so what you're getting on your phone, what you're hearing on TV is not just like what Jane, your local forecaster, came up with.

1:01:16She's relying heavily on these central predictions from major resources. So, for example, if worldwide governments decide we don't need these computers anymore, we don't need these satellites, it's not like you can be like, that's cool. I got my local weather forecaster. I don't need you. No, your local weather forecaster is getting that information from these big models that are being run by the government. Oh, wow. Yeah. And did NOAA get cuts recently? I'm going to bet they did. There were some talk about cuts. I don't know how much of that is going through. It's all kind of scary. It's hard to know.

1:01:50Yeah. Okay. All right. We won't get into that. Moving on. But amazingly, currently, we can pretty accurately predict the weather five to six days in the future. And you mostly remember when the weather prediction is wrong. You mostly don't realize that most of the time it's right. It tells you it's going to rain. It tells you it's going to be sunny. It's mostly correct. It's amazing. But there's still challenges. Things are not perfect. One of the biggest challenge is just incomplete information. You know, we don't have sensors in enough places and we don't have enough sensors. And sometimes this data availability changes.

1:02:25You know, things go offline or come online. Now your model has to compensate for that. I don't have the data. Do I assume it's similar to the past? Do I try to ignore that kind of data? It's not easy to be running a model if the inputs are constantly changing. The bucket had a hole in it. So now you don't have good bucket data. Exactly. Exactly. They got a new kind of bucket. You don't know how to calibrate it. I spoke to John Martin. He's a professor of meteorology. And he said that this might be the biggest challenge is how to combine the data to make a high quality initial state. That's one challenge.

1:02:58The other are these subgrid parametrizations. Can we develop better models for turbulent flow at the boundaries or for latent heat release back into the environment? And another limiting factor is just the computing cost. More computes, more GPUs from NVIDIA means smaller grids, which means the effect of these approximations, these parametrizations is less. Another continuing challenge are rare and extreme events. We're pretty good at predicting the bigger picture. Is it going to be sunny here? Is it going to be rainy here? But small, rare, extreme events, like there's a tornado right here, that's more challenging because they depend in detail on things that happen within the grid that we're averaging over.

1:03:41And so there's a lot of work being done right now. One thing we're hoping to do is like, let's reduce the grid size, get more computing, more accurate, right? But another really promising era of research is using machine learning. There's this movement in many fields of science to use machine learning to make predictions by essentially skipping the physics. Like the physics is hard. It takes a lot of time to push the initial conditions through these equations. In the end, you have an input and an output. And the idea is, well, can we train machine learning, not a chatbot, not LLMs, right? It's AI, but it's not LLMs, to map the initial conditions to the output.

1:04:22Because in the end, it's just a mapping and one could learn it. And so we have these machine learning models that are simple functions that take the input and give you the output. And they don't have the physics encoded in them, but they learn from the simulations. They learn the patterns. They learn what the rules are implicitly. And so you don't have to go through all the detailed calculations. So this can dramatically speed up your predictions. We use these at the Large Hadron Collider all the time so that we don't have to, for example, model every single particle that might hit the detector and create another particle and another particle.

1:04:55We can learn to predict the final thing we're interested in and just sort of leapfrog over all the tiny details. And is machine learning being used right now for weather predictions or are they just starting to work on how you would do that? They're using that now. They're sort of experimental, but there's a guy here at UC Irvine, Mike Pritchard, who is an expert in this kind of stuff. And it's very powerful. Absolutely. Cool. Yeah. So I asked John Martin, if I gave you a billion dollars to improve weather predictions, what would you do? And he said he would spend a billion dollars on ocean probes.

1:05:25Like he wanted a more substantial understanding of how water is circulating in the ocean and temperature in the ocean and how that's all working. because his suspicion was like, we're right next to this other big fluid that's affecting our temperature and we don't have much enough data about it. We just knew more about the ocean. And this just highlights like how little information we have. It's not just a question of like puzzling out the rules of the universe, but just like knowing what's happening. If we had more data everywhere about temperature, pressure, about cosmic rays, we would just learn so much about the universe.

1:05:58And we have so few ways to probe it. We're really just like taking the tiniest teaspoon out of this massive river of data and trying to use that to understand the whole river. It's crazy. How good do you think weather prediction would have to be before people stopped complaining about weather prediction? I asked John that question and his prediction was, quote, the complaining will never stop. Amazing. I think that, you know, weather prediction has improved a lot over the last few decades. It used to be you couldn't get any reliable prediction more than a day in advance. Now, five, six days is pretty reliable.

1:06:31But people expect that and they get used to it. And they're like, what? You didn't predict the weather on my ski trip in two weeks? I'm mad at you. And so, yeah, the complaining will never stop because we always just get used to the level of technological prowess that we've had. And so people want more because it's so important. And it's a hard problem. There's so much physics here. There's instrumental science. There's so many different kinds of science and interface with each other. It's really an exciting field. And let me throw a special thanks to Professor Jane Baldwin here at UCI, who told me a lot about weather predictions, and Professor John Martin at Wisconsin, who answered a lot of naive questions of mine.

1:07:07Thanks to both of you. Thank you, community. All right. See you all next time. I hope the weather is nice where you are. It always will be nice where I am.

1:07:23Daniel and Kelly's Extraordinary Universe is produced by iHeartRadio. We would love to hear from you. We really would. We want to know what questions you have about this extraordinary universe. We want to know your thoughts on recent shows, suggestions for future shows. If you contact us, we will get back to you. We really mean it. We answer every message. Email us at questions at danielandkelly.org. Or you can find us on social media. We have accounts on X, Instagram, Blue Sky, and on all of those platforms, you can find us at D &K Universe. Don't be shy. Write to us. Introducing the all-new Mazda CX-5.

1:08:05Featuring more connection. Hey, Google, where's the nearest Pilates class? Safety that has your back. More discovery on the scenic routes. More passion in the details. And more control in changing weather. The all-new Mazda CX-5. More to move every side of you. See it in five films at mazdausa.com slash five sides. Google is a trademark of Google LLC. Sequences shortened and simulated. This July 4th, come celebrate at America's Block Party, hosted by America 250. America's Block Party is a can't-miss 4th of July concert, happening at the Los Angeles Memorial Coliseum. Experience music performances by major artists, patriotic tributes and the kickoff to giving forth helping to make july 4th the largest day of giving in american history it's more than just fireworks join this landmark celebration and get your america's block party tickets now for 17.76 at america 250.org slash la running a business shouldn't feel like surviving a software group project one app for accounting another for inventory another for sales and somehow none of them talk to each other that's That's where Odoo comes in.

1:09:15An all-in-one business management software that brings every part of your business together. From sales and accounting to inventory and marketing. All in one powerful platform. No messy integrations. No bouncing between tabs. And best of all, no spreadsheets. Stop managing software and start managing your business with one unified system. Try for free today at odoo.com slash iHeartRadio. That's O-D-O-O-O dot com slash iHeartRadio. Here's the truth. You could literally be adored by everyone and then come home and still get completely ignored by your own cat. It's classic cat behavior. But new Shiba Premium Puree is a lickable treat that changes all that.

1:09:57They're protein-rich, made with bone broth, and have the smooth, creamy texture cats go crazy for, especially when it's hand-fed. Yeah, it's more than a treat. It's a fast pass to favorite human status. So feed your cat Sheba and go from totally ignored to truly adored in just 12 days, guaranteed, or your money back. Learn more at Sheba.com. This is an iHeart Podcast. Guaranteed human.