Artificial Intelligence (AI) has permeated the technological world over the last several years, with recent migration into the world of Numerical Weather Prediction (NWP).  Numerical Weather Prediction is a text-book reference to “Weather Prediction.”  There are many questions regarding AI’s strengths and weaknesses compared to the regular, conventional NWP models that have provided weather forecasts for decades.  

While the approach to generating an AI weather forecast is distinctly different from the approach completed by NWP, the two share similar fundamental steps in generating a weather forecast.  Both require an initial forecast, which tells the model, “This is the current state of the atmosphere.”  NWP and AI equations are also both fine-tuned using historic weather observations.  Lastly, both types of prediction are iterative. Iterative means that they complete the same forecast process for each step of the forecast using their own generated forecast output as input for the next iterative step in the forecast process.    

Where the two forecast methods differ from each other is the type of equations used to make an atmospheric forecast prediction. AI weather forecast models create their own forecast equations, whereas NWP models utilize the governing atmospheric equations which are founded on the laws of physics.  By creating its own equations, AI-driven forecasts are not nearly as computationally demanding as the physics-based equations that drive NWP.  This provides a distinct advantage to AI-based forecasts, especially AI models that generate accurate forecasts in the short-term (0-48 hours) when high-impact systems may be unfolding across a given forecast area.  However, while AI can generate a forecast much faster, the AI model does not use the laws of physics.  This limits a meteorologist’s ability to determine why the AI-based forecast was not accurate.  Because a meteorologist knows and understands the equations in an NWP model, the meteorologist can analyze a forecast and determine to much more confident level why the forecast generated the way it did. 

So what exactly is AI? AI is a scientific, technological field that combines highly detailed datasets (such as weather and climate data) and the world of computer science.  The result is a computer program that can systematically solve complex problems.  AI includes the sub-fields of Machine Learning (ML), which requires a human-labeled dataset to train the program (Supervised Learning), and Deep Learning (DL), which can use Supervised Learning or attempt to classify data on its own (Unsupervised Learning).  ML and DL are two different approaches that seek to organize historic weather and climate data to explain the “why” behind weather unfolding as it did. Once the AI-model has identified (via the data organization and analysis) why it “thinks” a weather event happened the way it did, the equation developed by the AI model is applied to an initial forecast state to generate a future forecast.

AI-driven weather forecast models show strong achievement in reducing the amount of time required to create a weather forecast.  However, like NWP, AI-driven models struggle with forecast accuracy at longer lead times.  This is a result of AI-based forecast models “training” on NWP weather forecast output.  Thus, overall, AI forecasts are not better than NWP forecast accuracy, and it is suggested that AI weather forecast models be used as a forecast tool in conjunction with NWP, not a stand-alone resource for weather forecasts. 

Climate Impact Company uses AI-generated forecasts primarily for medium-range (6-10-day/11-15-day) time scales. These forecasts are used to determine weather pattern change which is essential to energy and agriculture commodity markets. We began using AI forecasts steadily during mid-summer. During the past 30 days, the Artificial Intelligence Forecast System (AIFS) has a warm forecast bias in the Midwest to Southeast U.S. for the 6-10-day period although not nearly as warm as produced by the highly relied upon “European” model.