Research Interests

My work centers on advancing the understanding of atmospheric chemistry and its intersections with human and environmental systems. I am particularly interested in leveraging innovative data science and machine learning techniques to address complex challenges in air quality and wildfire land/smoke management. I enjoy approaching problems from unique perspectives, uncovering insights through the exploration of unconventional methods and data sources.


Current Research Areas
Wildfire emissions
Wildfire emissions contributing to smoke exposures across the western US in September 2020. Black dots are prescribed burns greater than 1000 acres during 2015–2020.

Wildfire, Prescribed Fire Smoke Modeling and Mitigation

Due to a warming climate, a legacy of fire suppression, and expanding development into the wildland-urban interface (WUI), the western US has experienced a recent rise in extreme wildfire seasons. Wildfires not only damage ecosystems and infrastructure but also degrade air quality and pose serious public health risks from smoke exposure. Prescribed fire is often promoted as a policy solution in the western US, yet its use remains limited in practice, and few studies have evaluated its effectiveness during recent wildfire seasons. My research is motivated by key gaps in our understanding: (1) we lack observational and modeling systems to accurately project how large-scale prescribed fire use would affect air quality and health outcomes in the western US; (2) the efficacy of past prescribed fire treatments remains poorly quantified across varied landscapes and fire seasons; and (3) it remains unclear whether expanding prescribed burning will reduce wildfire risk or simply add to the smoke burden without preventing future fires. Some of my recent work shows that prescribed fire, while modestly effective, are frequently least successful in the WUI, a central focus of wildfire policy. Such findings highlight the limitations of current wildfire strategies and underscore the need for data-driven, policy-relevant approaches to guide the proposed expansion of prescribed fire.


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

Co-Authored Wildfire Publications: Qiu et al., (2024), Liu et al., (2024)


Machine-learned solver
Machine-learned chemical solver embedded in the 3-D chemical transport model GEOS-Chem.

Deep Learning Atmospheric Chemistry

Global models of atmospheric chemistry are computationally expensive. The chemical solver that integrates the large-dimensional coupled systems of kinetic equations describing the chemical mechanism is a bottleneck. Machine learning (ML) could be transformative for reducing the cost of an atmospheric chemistry simulation by replacing the chemical solver with a faster emulator. My past work found that ML chemical solvers experience rapid error growth and become unstable over time. This challenge was addressed by introducing physical constraints into the ML architecture and training the ML solver online synchronously within the global atmospheric chemistry model. This approach enabled, for the first time, full-year global simulations of air quality using an embedded neural network solver. However, challenges remain in extending ML solvers to more complex chemical mechanisms and in reducing errors associated with long-term chemical aging. My ongoing research focuses on addressing these limitations by developing transformer-based architectures for improved spatial/temporal generalization, applying transfer learning from atmospheric AI foundation models, and leveraging self-supervised training algorithms specialized for remote sensing data. These efforts ultimately aim to develop accurate, stable, and physically consistent ML models to enable the simulation of comprehensive atmospheric chemistry in climate and Earth System models.


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

Co-Authored ML Publications: Balasus et al., (2023)


PM2.5 sensor locations
Distribution of sensor locations in the EPA monitoring network compared to those identified as optimal by the compressed sensing (mrDMD) algorithm in the western US.

Data-Driven Air Pollution Sensing

Despite major investments in air quality (AQ) monitoring, existing sensor networks often fail to capture extreme air pollution. My research uses data-driven methods to improve the design of sensor networks, air quality forecasts, and environmental early warning systems. In one national-scale study, I applied compressed sensing algorithms, a signal processing method that uncovers key spatiotemporal patterns found in data, to determine optimal AQ sensor locations based on recent pollution trends. This analysis revealed major gaps in the current EPA's monitoring network across the western US, particularly in regions affected by wildfire smoke. In a related study, I incorporated equity constraints into the sensor network optimization to improve coverage in historically under-monitored neighborhoods in cities such as St. Louis and Houston. These approaches provide a foundation for rethinking how we design air quality monitoring networks to better capture extreme events and ensure more equitable coverage. At the same time, commercial platforms increasingly deliver AQ forecasts through proprietary systems, raising concerns about transparency and public accessibility. These systems are likely to become more prevalent in the coming decade due to the rapid commercialization of environmental data and advances in AI and cloud computing. In response, my research is guided by a set of core questions: What are the early warning signals of extreme air pollution? Can open data outperform commercial forecasts? And can we learn more from air quality measurements than what is directly observed?


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

Co-Authored Sensor Publications: Kawano et al., (2025), Yang et al., (2022)

Past Projects
Chemical data assimilation figure
Interpolated contour plots of ozone concentrations as a function of latitude and altitude in April-May 2018. Observations from the ATom-4 aircraft campaign are compared with the Control and Full Assimilation simulations in GEOS-CF sampled along the aircraft flight tracks.

Chemical Data Assimilation for Atmospheric Composition

The NASA Goddard Earth Observing System Composition Forecast system (GEOS-CF) provides global near-real-time analyses and forecasts of atmospheric composition. The current version of GEOS-CF builds on the GEOS general circulation model with Forward Processing assimilation of meteorological data (GEOS-FP) and includes detailed GEOS-Chem tropospheric and stratospheric chemistry. Here we add 3D variational data assimilation in GEOS-CF to assimilate satellite observations of ozone including MLS vertical profiles, OMI total columns, and AIRS and IASI hyperspectral 9.6 μm radiances. We find that the detailed tropospheric chemistry in GEOS-CF significantly improves the simulated background ozone fields relative to previous versions of the GEOS model, allowing for specification of smaller background errors in assimilation and resulting in smaller assimilation increments to correct the simulated ozone. Comparisons to independent ozonesonde and aircraft (ATom-4) observations for 2018 show significant GEOS-CF improvement from the assimilation, particularly in the extratropical upper troposphere.


Publications: Kelp et al., (2023) ERL


Vehicle emissions figure
Road and street classifications in Los Angeles County with county base map provided by the U.S. Census Bureau's Master Address File / Topologically Integrated Geographic Encoding and Referencing (MAF/TIGER) Database (MTDB). (src credit: Yurika Harada)

Vehicle Emission Factors for Area-Wide Mobile Monitoring

On-road vehicle emissions are a significant source of outdoor air pollution, which pose a severe human health risk, especially for those who live near busy roads. A city's vehicle fleet can determine the levels of risk and exposure for residences as emissions differ for gasoline- vs. diesel-powered vehicles. Most research sampling mobile source emissions in urban traffic involves “vehicle chase” studies of exhaust plumes from individual vehicles which may not be representative of the average emissions for a given area.

We create a statistical model from measurements obtained from continuously moving platforms to estimate area-wide average vehicle emission factors of neighborhoods. These model predictions are used to estimate emission factors by source-related features within a city. Furthermore, our model can calculate separately light-duty and heavy-duty vehicle emission factors for a study area while also separating out high-emitter vehicles that may artificially skew emission factor estimates. Study areas include Los Angeles, USA and Chengdu, China.


Publications: Kelp et al., (2020) Atmos. Env.

Related Publications: Wen et al., (2019)

Cookstove intervention figure
Diurnal distribution of real-time PM2.5 concentrations with 10-min resolution. “S1” (A) is the pre-intervention baseline and “S2” (B) is the post-intervention follow-up using cookstoves.

Indoor Air Pollution from Cookstove Interventions in S. India

Biomass combustion from residential cookstoves is a major source of indoor air pollution and a large contributor to the global burden of disease. Investment of resources into rural energy intervention programs has great potential to improve household air quality in developing countries and thus increase quality of life and improve public health. We conducted a randomized intervention study to evaluate air pollution impacts of a carbon-finance-approved cookstove in rural South India. We employed real-time monitors to measure indoor concentrations of PM2.5, black carbon (BC) and carbon monoxide (CO) in households using carbon-finance-approved stoves and households continuing to use traditional open fire stoves. Implementation of the new cookstoves decreased concentrations of CO and PM2.5, but increased BC concentrations relative to the traditional stoves.

Although lab studies have clearly demonstrated the potential benefits of cookstove interventions, achieving these same benefits in real households is more complex. This work suggests that reduction in indoor pollution from intervention cookstoves might not be occurring in practice to the same extent as is expected from lab evaluation, and that benefits from such interventions should not be assumed.


Publications: Kelp et al., (2018)