GeoAI Unpacked #2: AI for weather forecasting
Welcome to GeoAI Unpacked! I am Ali Ahmadalipour and in this newsletter, I’ll be sharing insights and deep dives in geospatial AI, focusing on business opportunities and industry challenges. Each issue will highlight key advances, real-world applications, and my personal perspective on emerging opportunities. For a sneak peak at the topics I plan to cover in upcoming issues, visit here.
This issue explores the topic of AI for weather forecasting. I put a lot of time and effort into researching and writing this article. If you found it valuable, please share it to help spread the word.
1. Introduction
Weather forecasting has always played a critical role in safeguarding lives, driving economic activity, and enhancing day-to-day decision-making. Accurate predictions help many different sectors, and the ability to foresee and prepare for weather changes has long been a cornerstone of societal resilience.
Historically, weather forecasting has been one of the largest consumers of computational power. Advances in supercomputing enabled meteorologists to process vast amounts of atmospheric data, leading to more accurate forecasts. Today, the revolution brought by artificial intelligence (AI) is transforming the field once again. Machine learning algorithms are enabling faster, more precise predictions by analyzing complex datasets, opening new doors for both public and private sectors to deliver hyper-local and tailored forecasts.
2. Evolution of weather forecasting
Long before logistic regressions were hailed as "ground-breaking AI technology" in some industries, and startups began raising millions in VC funding for "revolutionary" methods in sectors they barely understood, statistical methods and AI were already widely used in weather forecasting. The figure below provides an overview of the evolution of weather forecasting, highlighting key milestones and the overall resolution of models across different eras. It specifically focuses on public weather forecasts generated by national weather agencies like NOAA and ECMWF, commonly known as the US and European models, respectively.
These advancements have substantially improved global weather forecasts, as demonstrated by Bauer et al. (2015), and have reduced forecast errors for extreme event forecasts, such as tropical cyclone track predictions, as shown in the figure below.
While data assimilation and various post-processing techniques have proven to be significantly effective for weather forecasting, the advances in weather forecasting during the past 3 years have primarily come from data driven AI-based models. Therefore, let’s look at the details of such in the next section.
3. AI-based weather forecasting in a nutshell
Traditional Numerical Weather Prediction (NWP) uses physical models and empirical functions to simulate the atmosphere, solving mathematical equations based on physics and observations. These models solve complex mathematical equations based on fundamental laws of physics (e.g. fluid dynamics, thermodynamics, and radiative transfer). NWP models simulate how variables like temperature and wind evolve over time, and they are computationally intensive and can struggle with small-scale or long-term predictions due to uncertainties.
AI-based weather forecasting, on the other hand, leverages machine learning models trained on past weather data to predict weather patterns directly from data without explicitly solving physical equations. They are typically faster, can handle vast amounts of data, and may excel in identifying non-linear relationships that traditional NWP models might miss. However, AI models often require large amounts of high-quality data for training and might not generalize well to situations outside their training range.
The summer of 2023 marked a turning point for AI-based weather forecasting, shifting the narrative from “Should we trust AI to predict natural disasters?” to “How Big Tech AI models nailed forecast for Hurricane Lee a week in advance”. Over the past couple of years, many leading institutions have primarily focused their weather forecasting R&D efforts on AI-based approaches.
Several AI methods including Graph Neural Networks, Vision Transformers, Foundational models, and hybrid ML+physics models have been employed for AI-based weather prediction. Numerous research groups are actively engaged in AI-based weather forecasting, continuously developing innovative methods. This field represents a vibrant and dynamic area of research, with ongoing advancements and discoveries. Amazingly, anyone can now run AI-based weather models using ECMWF’s open data!
Sophisticated mimickers or physics interpreters?
Several studies have assessed the generalizability of AI-based weather models to see if they are merely sophisticated mimickers (i.e. through pattern matching and interpolation) or can they encode the underlying physics too. The general consensus is that AI-based weather models are able to predict dynamic structures never seen during training periods and the local disturbances evolve and propagate in their results in a physically meaningful manner. Notably, there’s a limit to such predictability and how unprecedented and outlier an event can be. AI-based weather models generally fail to produce meaningful results if they are provided with several degrees warmer climate and oceanic conditions that may occur in distant future as a result of climate change.
4. Private weather forecasting ecosystem
Private weather forecasting has existed for over six decades, with numerous small and large companies operating in this sector. The business model has remained largely consistent: offering specialized weather services to industries such as agriculture, aviation, and energy, which require more tailored forecasts than those provided by government agencies.
Here, I categorize the private weather forecasting ecosystem into four groups: traditional forecasting companies, startups focused on AI-based weather predictions, weather teams within big tech companies, and research centers (even though they aren’t necessarily private institutions).
4.1. Traditional weather forecasting companies
In this category, companies typically take public weather forecast data and apply post-processing techniques to enhance forecast accuracy. These techniques range from statistical bias correction and downscaling (to capture finer regional details) to advanced machine learning methods. Some companies have even implemented physics-based models like WRF to run their own forecasts in-house, enabling them to customize data assimilation and parameterization. A few companies have gone further, launching their own constellations of satellites or radars to improve the initial condition assessment (though this is a significant challenge).
📝 Example of companies in this category are (in alphabetical order):
AccuWeather, Atmospheric G2, ClimaVision, DTN, Meteomatics, MyRadar, Spire, Tomorrow.io, WeatherOptics, and Weather Stream
4.2. Data-driven (AI-based) weather companies
As noted earlier, the data-driven approach gained traction only a couple of years ago, which is why most companies in this category are startups or recent entrants. Notably, some of the companies in the traditional category are also exploring this area, particularly as many AI-based models have become open source and more accessible. RWE, one of the world’s largest renewable energy developers, is a prime example of a major company that has assembled a talented team of meteorologists and researchers focused on developing AI-based weather models.
📝 Example of startups in this category are (in alphabetical order):
Atmo, Brightband, Excarta, Jua.ai, Silurian AI, Vaisala Xweather, Windborne Systems, Worldsphere.ai, and Zeus AI
4.3. Weather forecasting at big tech
Companies like Amazon and Apple have had weather teams for a relatively long time to support their operations. Those teams are primarily focused on logistics and operations rather than research. In 2015, IBM acquired The Weather Company for $2 billion, one of the first major acquisitions of a weather business. However, they sold it in 2023 for about half that value.
The progress of deep learning and availability of structured weather datasets (i.e. ERA5) allowed various people to look at weather forecasting as a data driven machine learning problem. Huawei released the PanGu weather model in 2022 and that got a lot of attention. Meanwhile Microsoft, Google, and NVIDIA all have dedicated teams that are actively working on weather (and climate) forecasting.
Although many researchers are working in these teams at big tech companies and occasionally publishing fascinating studies, it's uncertain whether these efforts will develop into a major business for those firms. So far, they seem more like showcases of AI research or ways to claim breakthroughs in public relations rather than driving substantial business outcomes.
4.4. Research centers
Several research centers are at the forefront of developing AI-based weather forecasts. ECMWF, a leading institution in this field, has significantly expanded its AI research team in recent years. Their WeatherGenerator project aims to revolutionize Earth system modeling using AI. In the US, the NSF AI Institute for Research on Trustworthy AI in Weather, Climate, and Coastal Oceanography (AI2ES) is a collaborative effort focusing on AI applications in these areas. The Allen Institute for AI (Ai2) is another notable contributor, conducting innovative research on AI-based weather and climate modeling.
5. Advantages of AI-based weather forecasting
5.1. Light-weight inference
Training data-driven weather models demands significant data and computing power, but running them operationally (inference) is much faster and requires far less compute. Due to their low compute needs during operational runs, AI-based models can generate super-ensembles, allowing for 1,000 probabilistic runs instead of the 50 to 100 typically produced by current operational models.
5.2. Potential for scientific discoveries
AI's capability to uncover nonlinear dynamics and complex interactions in the atmosphere suggests it could excel in modeling various atmospheric processes. There's also potential for extending the current limits of weather forecast predictability, as illustrated in the following figure from the Washington Post:
A recent study using the GraphCast AI weather model highlighted the importance of initial conditions, and it showed that it’s possible to achieve a 90% reduction in 10-day forecast errors for the 2021 Pacific Northwest heatwave if the right initial condition is employed.
5.3. Replacing outdated legacy codes
Many weather forecast models used by national weather services today rely on legacy code developed years or even decades ago. These codes were primarily written in languages like Fortran or NCL (e.g. HRRR), which have become outdated or too specialized, leaving few experts capable of maintaining them. Even the output formats, such as grib files, are outdated and cumbersome to work with. As a result, improving these legacy models will soon become nearly impossible.
AI-based models offer a fresh start, enabling the development of code using modern tools that are more accessible and easier to maintain for a wider audience.
5.4. Long-term potential (maybe)
A key challenge with climate change is that global climate models (GCMs)–that are currently used to project future changes–are coarse in resolution and rely on simplistic assumptions or ignore some atmospheric processes entirely. This makes them unable to accurately model precipitation or local phenomena like tornadoes and storms. AI-based models offer hope for better understanding and simulating these processes with less bias and greater accuracy. A recent study suggests that these models could significantly contribute to climate science. However, there remains a lack of consensus on their capabilities, with some researchers expressing concerns and reservations about their limitations.
Ideally, AI-based models would take potential future oceanic and atmospheric conditions to generate ensembles of likely hurricane paths in the future, for instance. This could help assess risks, guide adaptation strategies, and accurately price the associated risks.
6. Challenges of AI-based weather forecasting
6.1. The challenging business of private weather forecasting
This isn’t an issue exclusive to AI-based models but reflects a broader challenge within the private weather forecasting industry. These companies often make for strong consulting businesses, but they tend not to scale easily. You can run a successful profitable private weather forecasting firm with 10-30 employees, but expanding to 500 employees with massive revenues is difficult. Here are a few reasons why:
6.1.1. Free vs paid forecasts
Private weather forecast companies face tough competition from FREE forecasts provided by government weather agencies, which are often sufficient for many use cases and perform well in predicting many extreme events. For example, the following timelapse shows the National Hurricane Center's track forecast for Hurricane Milton, the second-most intense Atlantic hurricane recorded in the Gulf of Mexico. Remarkably, their first advisory—issued 4 days before landfall, when the cyclone hadn’t yet formed—predicted landfall just 12 miles north of the actual landfall location.
6.1.2. Data vs. solution
Weather data sits at the bottom of the value chain, meaning forecast data alone generates low revenue since people aren’t willing to pay much for it. This creates a major challenge: achieving high revenues requires acquiring a large number of customers in a highly commoditized market, where long-term customer retention can be difficult. On the other hand, weather solutions tailored to specific industries (e.g., renewable energy) can generate more revenue, but often require customized tools that may not be scalable across other clients, leading to a consulting-heavy model.
A similar challenge exists in the Earth Observation industry, where satellite data providers struggle to build solutions without competing with their own customers and product cannibalization.
6.1.3. Wealthy clients value exclusivity
Hedge funds, particularly in commodity and energy trading, are keen on private weather forecasts, as even slightly better accuracy can significantly boost profits. However, they seek unique insights; if everyone has access to the same information, it erodes their competitive advantage. This makes catering to these clients somewhat of an oxymoron in relation to scaling efforts. Additionally, major hedge funds employ their own meteorologists and developers who focus on R&D for advanced forecasting techniques, including AI-driven weather models.
6.2. Reanalysis vs. observation
Many AI-based weather models rely heavily on reanalysis datasets like ERA5, which often struggle to accurately represent real-world weather conditions, particularly for variables such as precipitation and wind speed. These discrepancies arise due to factors like topography, urbanization, and vegetation, leading to low correlations between ERA5 and ground-based observations. Notably, ERA5 temperature data can exhibit significant biases. While regional studies have explored using observation data for AI-weather modeling (here’s an example), global-scale observation-based models have been less successful due to the limited availability of observational data compared to reanalysis products.
Consequently, most AI-weather models still require the output of traditional numerical weather prediction (NWP) models like ERA5 to generate forecasts. However, research teams are actively working to address this limitation, and the development of purely observation-based AI-weather models holds the promise of significant advancements in the field.
6.3. Difficult to model precipitation
AI-based weather models often struggle with precipitation forecasting due to several factors such as:
Complexity of precipitation processes
Nonlinear relationships and complex interplay of atmospheric variables
Regional-scale variability of precipitation rate
Data quality and availability
Sparse and inconsistent observations
Measurement challenges (e.g. instruments failing due to storms or wind)
Sensitivity to initial conditions
Model limitations
Insufficient training data
Model architecture and objective function
Ongoing research seeks to overcome the above challenges hindering the accuracy of AI-based precipitation predictions.
6.5. Interpretability and trust
Similar to many other AI applications, AI weather models can produce outputs that may appear unusual or difficult to interpret. These outputs may exhibit spatial inconsistencies or unrealistic patterns, even if the overall error metrics are favorable.
For example, AI weather models may generate results that seem 'blurry' or 'smeared out' compared to traditional NWP models. Consider the example shown in the following figure: the left plot shows ECMWF's NWP model, while the right shows their AI weather model (AIFS). Both depict 2m temperature anomalies for a 48-hour lead time. While the overall patterns are similar, the NWP model appears sharper and more realistic, while the AIFS output is blurrier. This raises the question: do we prioritize 'inaccurate but visually appealing' forecasts, or 'inaccurate but blurry' ones?
The blurriness often observed in AI weather model outputs can be attributed, in part, to the objective function used during training. Minimizing the Root Mean Square Error (RMSE) can inadvertently lead to such artifacts. Recent research has explored the use of diffusion models and generative AI techniques to address this issue. By modeling correlated noise, these approaches have shown promise in generating sharper and more realistic forecasts.
7. Final remarks
This article explored the applications of AI in weather forecasting, with a particular emphasis on AI-based weather models. Given the rapid pace of advancements in this field, I have tried to include as many relevant links and resources throughout the text.
To conclude, I would like to briefly touch on a few additional points:
7.1. Weather is not climate
This article primarily focused on weather forecasting, which typically extends up to two weeks into the future. There are a lot of startups that are developing AI applications for sub-seasonal to seasonal (S2S) climate predictions, but I chose not to include them in this discussion due to their different focus. I’ll hopefully explore that subject more thoroughly in the future.
📝 A few startups for S2S climate forecasting are (in alphabetical order):
Beyond Weather, ClimateAi, Planette AI, and Salient
7.2. On the importance of benchmarking
The saying "You can't improve what you can't measure" is particularly relevant in weather forecasting. The vast number of weather stations worldwide allows us to quickly evaluate forecast accuracy for many locations. However, gaps in coverage, especially in remote areas and the upper atmosphere, necessitate the use of reanalysis products like ERA5. WeatherBench's initiative to compare various AI-based weather models against ERA5 and IFS HRES t0-analysis is a valuable contribution to the field. By increasing the number of such third-party evaluations (especially the regional ones), we can more accurately identify model shortcomings and foster improvement.
That said, using ERA5 as a benchmark presents a challenge because many AI models are trained on it. Optimizing to minimize error against ERA5 limits the AI model's potential, making it no better than ERA5 itself (which has its own biases). Therefore, moving away from reanalysis products—at least for metrics like precipitation, near-surface air temperature, and wind speed where ground observations are available—is crucial for maximizing the operational performance of AI-based weather models.
7.3. Communication and evacuation
Despite significant advancements in weather forecasting, extreme events continue to pose substantial risks and result in tragic consequences. Hurricane preparedness and evacuation highlight this issue. For example, forecasts for Hurricane Milton were accurate 3-4 days ahead of landfall, but the road network couldn’t handle the evacuation of millions in that time frame. This led to traffic jams and uncertainty over whether people could evacuate in time. Similarly, during Hurricane Ida in 2021, New Orleans officials acknowledged that the city’s highway capacity requires evacuation orders to be issued at least 72 hours before tropical-storm winds arrive.
These cases underscore the critical need for not only accurate forecasts but also robust evacuation strategies and infrastructure that can effectively handle large-scale population movements during extreme weather events.
Outstanding summary.