Prescribed fire is widely proposed to reduce wildfire impacts, but its benefits for air quality and health depend on where treatments occur and how they affect later fires. Our research examines how prescribed fire use affects air quality and health outcomes, and how treatment effectiveness can vary across landscapes and fire seasons. A primary question is under what conditions expanding prescribed burning reduces overall smoke exposure after accounting for emissions from the burns themselves. Some of our recent work shows that prescribed fire treatments reduce burn severity and smoke emissions, but are less effective in the wildland-urban interface (WUI), where homes meet wildland vegetation.

Publications: Kelp et al. (2025) AGU Adv., Kelp et al. (2023) Earth's Future

Mentored Publications: Chung et al. (2025) ES&T,

Co-Authored Wildfire Publications: Li et al. (2026) Sci. Adv., Liu et al. (2026) ISPRS Arch, Feng et al. (2025) PNAS, Qiu et al. (2025) Nature, Qiu et al. (2024) ES&T, Liu et al. (2024) Int J Wildland Fire

Points for Policymakers: California's Fifth Climate Change Assessment Report (2026), Prescribed fire research brief (2025)

Ozone is an air pollutant that harms human health, damages vegetation, and acts as a short-lived climate forcer. Despite extensive observations, global atmospheric models still disagree on even basic ozone trends and spatial patterns. Accurately predicting ozone requires capturing nearly every major process in the atmosphere (e.g., emissions, chemistry, stratosphere-troposphere exchange, boundary layer mixing). Few problems in geoscience are as well-observed and yet as difficult to predict, which makes ozone a particularly revealing test for AI. Our past work explored how AI methods can be used to emulate and replace computationally costly components of global atmospheric models for fast, stable simulations. Our current research probes what AI foundation models learn about atmospheric chemistry and the extent to which they can complement (or even substitute) conventional process-based models. We use mechanistic interpretability to identify and test the internal representations driving AI model predictions with the goal of diagnosing errors and improving physical consistency.

Publications: Kelp et al. (2022) JAMES, Kelp et al. (2020) JGR: Atmos, Kelp et al. (2018) ArXiv

Future Priorities of AI in Air Quality: White Paper on Governance of AI Models in Earth Science (2026), Highlight Paper on AI for Tropospheric Ozone Research (2025), NeurIPS AI4Science Dataset Competition Winning Paper (2025)

Despite major investments in air quality (AQ) monitoring, existing sensor networks often fail to capture extreme air pollution. We use data-driven methods to improve the design of sensor networks, air quality forecasts, and environmental early warning systems. In a national-scale study, we used compressed sensing to determine optimal AQ sensor locations based on recent pollution trends. This analysis revealed major gaps in the EPA monitoring network across the western US, particularly in regions affected by wildfire smoke. In a related study, we incorporated equity constraints into the sensor network optimization to improve coverage in historically segregated neighborhoods in cities such as St. Louis and Houston. We also ask: What are the early warning signals of extreme air pollution (fires, inversions, smog)? Can open data outperform commercial forecasts?

Publications: Kelp et al. (2023) GeoHealth, Kelp et al. (2023) ERL, Kelp et al. (2022) ERL

Co-Authored Sensor Publications: Kawano et al. (2025) Sci. Adv., Yang et al. (2022) ES&T