AI leading the way for weather prediction
I believe in a very short timeframe (2–3 years), AI will be the sole input used for weather channels and resources to predict the weather. AI is already demonstrating a superior ability to predict the weather than the current “gold standard” supercomputer-based system current providers use to predict the weather, and many different companies with AI abilities (like Google) are working on an AI-powered weather predictor. In this post I will discuss current research on the topic, previous weather prediction models, and potential consequences and benefits of relying on AI for weather prediction. AI weather prediction produces predictions that are not only more accurate, but in a shorter time frame and are not as energy intensive to run.
Traditional weather prediction systems rely on mathematical models that simulate behavior of the Earth’s atmosphere, based on principles of physics and meteorology studies. These systems use spatial grid-based systems and conservation laws of mass, energy, and momentum to fuel the model equations for their prediction systems, and layer on top observational data from actual weather stations, satellites, radar and balloons to represent the current state of the atmosphere (“How do traditional weather predictors work?” prompt. ChatGPT, Open AI, 20 November 23). Some weather systems are focused on specific regions and sub-terrains, and some weather prediction systems are focused on larger zones and territories. The weather systems consider interactions with the Earth’s surface, surface temperatures, and current conditions to project likely weather patterns in future time frames, ranging from a few minutes, to hours, to days and even weeks. Naturally, the further out a prediction is from the current date/time, the higher likelihood of it changing (we as consumers have definitely experienced this!)
In ground-breaking AI technology, Google DeepMind recently created GraphCast, an AI weather forecaster, trained on 40 years of weather data using satellite images, weather stations, and radar systems, including current weather ranging from approximately 6 hours ago to predict the forecast roughly 6 hours in the future. According to a recent article published in Government Technology, GraphCast produces predictions for over 1 million grid points across the surface of the Earth and at 37 different altitudes in the atmosphere. Additionally, a core benefit of GraphCast is that it can run on a single TPU machine, versus the current weather machines that use multiple supercomputers and servers. In testing GraphCast’s abilities against current weather machines, it made 10-day weather predictions in under 1 minute that were more accurate than the supercomputer. (Government Technology, 15 November 2023. https://www.govtech.com/question-of-the-day/is-ai-better-at-predicting-the-weather-than-a-supercomputer). For further evidence of GraphCast’s efficacy, it predicted Hurricane Lee would make landfall in Nova Scotia 3 days earlier than the current weather prediction models said. GraphCast also can offer warnings much earlier than standard models of extreme temperatures and paths of cyclones. (Melissa Heikkila, Technology review, 14 November 2023. https://www.technologyreview.com/2023/11/14/1083366/google-deepminds-weather-ai-can-forecast-extreme-weather-quicker-and-more-accurately).
The consequences of AI surpassing accuracy of current weather predictors is that the weather industry on the internet, cable news, and in print (newspapers) may be either consolidated, go out of business, or need to drastically change work practices to incorporate AI into their daily weather updates. This may involve re-skilling, upskilling, or even potentially layoffs for the teams of meteorologists working on weather prediction now. This also has consequences for the larger global society, as the “weatherman” has been a synonymous symbol with the news for as long as we know. This will also radically change the entire world, because if weather prediction gets better, we will be able to better plan and prepare for natural disasters like hurricanes, tornadoes, earthquakes and more that cause destruction on communities who do not have time to evacuate or prepare. Weather prediction is one of the most difficult problems society has related to accuracy, and traditionally meteorologists use massive computer simulations with limited accuracy.
However, I believe it is important to raise one of the key constraints of using AI to predict the weather; AI uses previous patterns to predict future outcomes, however with climate change and rising global temperatures, it is likely that previous patterns may not be all that useful to predict future outcomes at the pace unprecedented natural disasters or weather patterns are emerging. This relates back to our in-class discussion around this exact limitation of AI in the self-driving car industry: it is nearly impossible to train self-driving cars on all of the possible tragic scenarios in order to avoid accidents, which is resulting in many autonomous cars getting in accidents or hitting people. We will need to design a way for AI to take into account the ‘human layer’ here — where meteorologists can update the inputs to these AI prediction models in order to take into account the unprecedented rate of climate change in order to have more accurate weather predictions on a daily or annual basis.
In conclusion, the weather industry is ripe for disruption from AI, largely for good of humanity. While these models are not perfect yet and still lag in some areas (for example, precipitation), in partnership with meteorologists, the capability of AI will quickly surpass the ability of current weather prediction models and will lead to more accurate and faster weather predictions, resulting in many benefits for society in the near-term.
