Brains
The National Oceanic and Atmospheric Administration recently launched three new artificial intelligence based weather forecast products. The A.I. Global Forecast System and the A.I. Global Ensemble Forecast System will now be routinely offered alongside its other weather model products. These new weather models offer slight increases in accuracy, but can be computed roughly 1,000 times faster than traditional numerical weather prediction models. NOAA is also offering a hybrid ensemble forecast, the HGEFS, which incorporates predictions from both the AIGEFS and the conventional GEFS. The new developments are the result of Project EAGLE, a collaboration between NOAA, the Earth Prediction Innovation Center, the Office of Oceanic and Atmospheric Research, and the National Weather Service. Project EAGLE aimed to standardize the process for gathering, formatting, and storing forecast and observational data for training new A.I. models, as well as streamline the process for operationalizing new advances into usable forecasts.
Researchers have been investigating the uses of machine learning for weather forecasting for about 20 years, but a lot of the excitement over operational A.I. weather forecasts came in November 2023 when Google DeepMind launched GraphCast. As the first large scale demonstration of huge computational speedup and moderate accuracy increase for a global weather model, GraphCast quickly inspired more interest from weather forecasting agencies like the European Center for Midrange Weather Forecasting. In January 2024, the ECMWF launched the AIFS, or Artificial Intelligence Forecasting System, which built on some of GraphCast’s methods, thanks in part to its open source release.
Training large machine learning models like GraphCast, the AIFS, or the AIGFS, which involves optimizing millions of parameters, requires huge amounts of data. When thinking about weather forecasting, there is no shortage of data. From remote weather stations like snotel sites to the many satellite observatories keeping an eye on the Earth, we’ve got plenty of data. However, one major hurdle to cramming all these observations and prior forecast data into a graph neural network is that these data are stored in many different formats on many different servers around the world. To surmount this hurdle, the ECMWF created Anemoi, named after the Greek gods of the winds. Anemoi is a framework for developing new A.I.-based forecast tools, streamlining data, training, model development, and validation against a standardized set of parameters. Project EAGLE, which led to the development of the AIGFS and AIGEFS, shares similar goals with Anemoi, but focuses on U.S. data and weather models.
Numerical weather models split the Earth and its atmosphere up into millions of boxes and solve the partial differential equations that physics gives us for the heat flow, moisture transport, and other properties to build up a model of what the Earth will look like one tiny step ahead in the future. Keep computing more and more tiny steps into the future, and you eventually get yourself to a 10-day forecast like that produced by the GFS. Solving partial differential equations billions or trillions of times each day is what takes up an immense amount of computing power. A.I.-based models skip a lot of the mathematics behind the physics of our atmosphere and just guess what the next state of the atmosphere will look like, based on all of the prior observations and forecasts the model was trained on. Skipping a lot of the math is where the 1,000 times speedup in computational time comes from, and the fact that these models can deliver similar accuracy to physics-based models like the GFS is remarkable.
Since the 10-day forecast is based on a series of steps, each one building on the predictions made in the previous step, the uncertainties in the forecast grow larger and larger as time goes on. Below is a plot of some of these forecast uncertainties for the GFS and several different A.I.-enhanced versions of the GFS, taken from a recent study from NOAA. For all of the quantities, be it temperature, wind, precipitation, or whatever else, the uncertainties grow linearly with time. 10 days out, these uncertainties blur the lines between a storm bringing in several feet of blower pow, a little bit of rain, or not showing up all together. Comparing the errors for the standard GFS and several flavors of GraphCast, it looks like machine learning does not offer much improvement in our ability to offer longer range forecasts, and the errors still grow with time. But, the tiny fraction of the computing time needed to produce these forecasts can offer substantial improvements in another area of forecasting.
The growth of forecast errors with time has plagued numerical weather prediction for as long as it has been around. Some of this error is due to the random nature of our atmosphere, which is, after all, just a lot of nitrogen, oxygen, water, and other components smeared over a rock hurtling through space. Another component of the error comes from our ability to measure the current state of the atmosphere, which has its own uncertainties associated with it. Ever compared a snow stake measurement to what you actually find on the runs at your local ski hill?
It turns out that forecasts can be extremely sensitive to the initial conditions from which they start. To overcome this sensitivity, ensemble forecasts incorporate slight variations in the initial conditions to see what impact they have on the forecast. The Global Ensemble Forecast System (GEFS) takes much of the same physics as the GFS and tweaks the initial conditions by a little bit to see what range of weather conditions we might expect. If it is slightly warmer, or if the winds shift slightly more to the north, do we still expect huge snowfalls? Below is a plot from the University of Utah’s Department of Atmospheric Sciences showing snowfall predictions for Big Sky, Montana, from an ensemble of forecasts for the upcoming weekend. The average looks to be right about four inches over the next 48 hours, but there are some ensemble members that predict more than eight inches or less than two inches, showing that at least for this storm setup, there is quite a lot of variation.
A.I. weather forecasts have the potential to improve ensemble forecasting systems by allowing a lot more ensemble members to be included. Computational time currently limits the number of initial conditions that can be routinely computed to one or two dozen, depending on the model. Researchers at Project EAGLE have talked about the potential for including up to 1,000 ensemble members, due to the speedup from using A.I.-based models. Boosting the size of the ensemble could lead to forecasts with more certainty, because wider ranges of initial conditions can be considered.
The impact that Project EAGLE and Anemoi will have on weather forecasting goes far beyond the new A.I. models that are currently operational. Building strong frameworks for model development and implementation will help accelerate the pace that new breakthroughs in both fundamental atmospheric science and machine learning will be able to improve our predictions of the weather. A focus on developing standardized datasets and toolkits will also broaden the field of people who may be able to contribute to new forecasting tools. The next few years are sure to bring new levels of accuracy, enabling better decision-making by ski area operators, more effective storm chasing, and more certainty that the next pow day will be worth taking that sick day.
More From Zach Armstrong:
- Antarctica’s Ice Sheet Is…Growing?
- East Coast Enjoys Historically Deep November While West Waits For Snow
- GoFundMe Helps Athletes and Para Athletes Make The Olympic Dream A Reality
- History with SnowBrains: The Incredible Notes from a Swiss Avalanche Researcher’s Travels to Western USA in 1949
- Sierra Nevada Resiliency Team Enters Second Season Supporting Ski Patroller Mental Health Across California and Nevada
