-0.65 parts per billion per year became +0.13. 3.9 years of clean-air progress effectively vanished. 43 million people were pushed into ozone nonattainment conditions during 2022-2024. Context second: a new Science study says wildfire emissions have helped flip the United States’ surface ozone trend from “slowly getting better” to “uh oh, the graph is climbing again” Deng et al., 2026.

On one hand, this is a triumph of measurement. Researchers used deep learning to build a daily, 1-kilometer surface ozone dataset for the continental U.S. from 2003 to 2024. That is like giving air pollution a surveillance camera grid, except instead of catching porch pirates, it catches invisible chemistry misbehaving in sunlight.

On the other hand, the camera found the house is on fire. Somewhat literally.

The Bad Ozone Is Down Here With Us

Ozone has branding issues. High up in the stratosphere, it is the noble UV shield we learned about in school. Down at lung level, it is photochemical smog, the atmospheric equivalent of a nice molecule joining the wrong group chat.

Ground-level ozone does not usually come straight out of a tailpipe or a tree. It forms when nitrogen oxides, volatile organic compounds, carbon monoxide, heat, and sunlight get together and start doing chemistry like unsupervised interns in a lab. EPA explains this basic recipe as pollutants reacting in sunlight to produce ozone near the ground, where people actually breathe it EPA.

For decades, the U.S. made real progress by cutting emissions from cars, power plants, and industrial sources. EPA’s 2025 air trends report says national 8-hour ozone concentrations dropped 24% since 1990, even while noting that fires have recently pushed up carbon monoxide, particle pollution, and ozone in some years EPA Our Nation’s Air.

That is the hopeful part. We made policy. The air improved. Civilization briefly appeared to know what it was doing.

Then the Fires Entered the Spreadsheet

Deng and colleagues argue that since about 2015, rising wildfire emissions have reversed the national policy-relevant ozone trend. Their model estimates ozone dropped by 0.65 ppb per year from 2003 to 2015, then rose by 0.13 ppb per year from 2015 to 2024. Remove fire impacts, they report, and the downward trend would have continued.

That is a nasty plot twist. We cleaned up a lot of the usual suspects, then climate-amplified fires walked in through the side door carrying carbon monoxide, nitrogen oxides, volatile organic compounds, and a very bad attitude.

The study also estimates that premature deaths from fire-sourced ozone have increased by 318 deaths per year since 2013, with post-2013 mortality 46% higher than before. In 2023, the Canadian wildfire season was so large that smoke-driven ozone helped expose huge parts of the U.S. to unhealthy conditions. On one hand, the atmosphere is an elegant connected system. On the other hand, “connected system” here means smoke from far away can show up like an uninvited cousin and ruin your local air-quality compliance plan.

Why Deep Learning Matters Here

The machine learning part is not decorative. Air monitors are useful but sparse, especially in rural areas. Satellites see broad patterns but do not directly measure every breath-level ozone value at every location every day. Deep learning lets researchers combine monitoring stations, satellite observations, meteorology, emissions, and spatial patterns into a “gapless” map.

Think of it as autocomplete for the atmosphere, except the stakes are lungs instead of accidentally texting “ducking.” The model fills in missing detail across space and time, then researchers ask: what changed, where, and how much of it came from fires?

That said, deep learning does not magically remove uncertainty. The model depends on training data, assumptions, exposure-response functions, and the messy chemistry of smoke plumes. Wildfire ozone can rise, fall, or move depending on sunlight, plume age, aerosol shading, and whether the smoke mixes with urban pollution. Recent work backs up that complexity: Lee and Jaffe found smoke days increased maximum daily 8-hour ozone across the continental U.S. from 2018-2023, while Hu and colleagues showed smoke particles can both suppress and enhance ozone chemistry depending on conditions Lee & Jaffe, 2024, Hu et al., 2025.

The Policy Problem Is Deeply Annoying

Ozone rules were built for controllable sources: cars, factories, power plants. But wildfire smoke does not respect county borders, state implementation plans, or the emotional needs of air-quality managers.

A city can reduce local traffic emissions and still get walloped by transported smoke. That makes accountability awkward. Who gets blamed when the pollution was cooked up hundreds or thousands of miles away? The atmosphere, tragically, does not attend jurisdictional meetings.

Other studies are pointing in the same direction. Cooper et al. found that Canada’s 2023 fires generated record-breaking ozone anomalies across the Upper Midwest Cooper et al., 2024. Chang et al. showed that U.S. surface ozone trends vary by season, region, and heat waves, which means climate pressure can complicate even successful emissions policy Chang et al., 2025.

On one hand, better satellites and deep learning give us sharper maps of the problem. On the other hand, sharper maps are a mixed blessing when they mostly help you locate the disaster with professional-grade precision.

The Part Where Hope and Dread Share a Bar Tab

This paper matters because it reframes wildfire smoke as more than a particle problem. PM2.5 gets the headlines because you can see the haze and feel it in your throat. Ozone is sneakier. The sky may look less apocalyptic, and your lungs may still be filing a complaint.

If the findings hold up, the real-world message is blunt: clean-air progress now depends not only on reducing tailpipe and industrial emissions, but also on managing fire risk, climate change, land management, forecasting, and monitoring. That is not a neat policy box. It is several boxes stacked badly in a garage.

Still, the research gives us something useful: a clearer view. And maybe that is where the mild existential crisis lands. We built models smart enough to reveal that yesterday’s environmental wins can be quietly repossessed by tomorrow’s climate-driven fires. I do not know if that makes me hopeful or terrified. Possibly both. Definitely both.

References

Deng, W., Wang, J., Zhou, M., Chen, X., Wu, X., Zhang, H., Cohen, J. B., Wei, J., da Silva, A., Brasseur, G. P., Granier, C., & Rouil, L. (2026). Fires reverse progress toward ozone air quality standards in the United States. Science. https://doi.org/10.1126/science.aed3197

Lee, H. J., & Jaffe, D. A. (2024). Wildfire impacts on O3 in the continental United States using PM2.5 and a generalized additive model (2018-2023). Environmental Science & Technology. https://doi.org/10.1021/acs.est.4c05870

Cooper, O. R., et al. (2024). Early season 2023 wildfires generated record-breaking surface ozone anomalies across the U.S. Upper Midwest. Geophysical Research Letters. https://doi.org/10.1029/2024GL111481

Chang, K.-L., McDonald, B. C., Harkins, C., & Cooper, O. R. (2025). Surface ozone trend variability across the United States and the impact of heat waves (1990-2023). Atmospheric Chemistry and Physics, 25, 5101-5132. https://doi.org/10.5194/acp-25-5101-2025

Hu, L., et al. (2025). Impacts of wildfire smoke aerosols on near-surface ozone photochemistry. Atmospheric Chemistry and Physics, 25, 8701-8718. https://doi.org/10.5194/acp-25-8701-2025

Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.