Quantitative Winter Ozone Forecasts

Project Updates

• Recent Developments:
  
  • https://basinwx.com/ 
  • We have written a method to score Clyfar's forecasts fairly. This matters because a forecast that warns too often gets ignored, like "the boy who cried wolf", while one that warns too rarely misses real events. We will post it publicly as a preprint (a paper shared before journal review) when is is complete. 
    
    
  • A paper that checks Clyfar's forecasts over the past five winters is in progress. It shows where and when Clyfar helped and where it did not, so this winter's forecasts can avoid the same mistakes.
    
    
  • John Lawson will present Clyfar and its AI-written forecast summaries at the American Meteorological Society Annual Meeting in Denver in January 2027, so forecasters across the country can learn from the Basin’s approach. 
    
    Current and Upcoming Work:
    
    
  • From 15 November 2026, a test version of basinwx.com will show short AI-written summaries of each Clyfar outlook (“Ffion”), in plain and technical versions. Ozone Alert emails will still be written by our forecaster. The test lets us check that the summaries are accurate and useful before anyone relies on them. 
    
    
  • The same test site adds improved live air-quality maps, so you can see current conditions across the Basin in one place. 
    
    
  • By 1 December 2026, two versions of Clyfar will run side by side to support Ozone Alert. Version 2 learns from where version 1 went wrong and uses machine learning to adjust its predictions and the guidance it gives Ozone Alert forecasters. Running both lets us check that version 2 really is better before we rely on it. 
    
    
  • New snow-depth stations (see Snow Cover Data Collection) will give Clyfar better snow information. Snow cover is one of the main ingredients of high winter ozone, so better snow data should mean more reliable warnings. 
    
    
    
• Problems:
  
  • Last winter had little snow, so we are re-running Clyfar over the past five winters to learn enough to improve it. This takes extra time, but versions 1 and 2 will both be running to support Ozone Alert by 1 December 2026. 

Example Clyfar model output showing Optimistic, Neutral, and Pessimistic represent the low, medium, and high range of forecast ozone values, respectively.

Improve Quantitative Winter Ozone Forecasts

Winter inversions are difficult to predict due to their inherent uncertainty, such as their sensitivity to snow depth, wind speeds, and temperature profiles. Small errors in the initial state of the atmosphere will become large errors in the coming forecast. Also, issuing solely “yes” or “no” forecasts for ozone events masks the true uncertainty of the prediction and can be misleadingly confident. In our team’s past attempts to forecast winter ozone episodes, our prediction methods compared well against historical datasets but performed poorly when we used them for real-time prediction (Lyman et al., 2020; Mansfield, 2018). Because of these failures, up to now we have more so  relied on qualitative (as in, descriptive or by experience) forecasts rather than numerical prediction methods for our Ozone Alert program.

To give the forecasting team more options to guide issued Alerts, we have constructed a fuzzy logic-based ozone prediction system named Clyfar that is thus performing better than past efforts (Lawson and Lyman, 2024), in part due to the addition of risk communication.  This system can predict high ozone events, including providing likelihoods of different ozone concentration thresholds, so users have more detailed information with which to make decisions.  We have also created a website to display products from Clyfar: https://basinwx.com.

In winter 2026–27 we will run two versions of Clyfar side by side for Ozone Alert, and use machine learning (a kind of artificial intelligence) to tune version 2 from the mistakes of past winters. We check every change against winters that were not used to build it, so an improvement is real and not luck. Outlooks and short AI-written summaries will appear on basinwx.com. Results will be published in peer-reviewed scientific journals, with plain-language versions in our fact sheets. We will also study the weather “tipping points” that make some ozone episodes hard to forecast, and keep asking forecast users what helps them make decisions, so the website stays useful to them.