Enabling High-Resolution Greenhouse Gas Flux Inversions with a Machine Learning Emulator of Atmospheric Transport

Abstract

Carbon dioxide (CO2) and methane are the two most important anthropogenic greenhouse gases (GHGs). Accurately quantifying GHG sources and sinks is essential for verifying emission reduction policies and projecting future climate trajectories. Recent works have shown that point sources dominate the GHG emission budget, and high spatiotemporal resolution is required to investigate their roles. Fortunately, there has been a proliferation of dense observing systems of GHGs, including surface-level and satellite measurements. However, traditional approaches of estimating GHG emissions from these measurements require running computationally and storage-intensive physics-based atmospheric transport models. These computational bottlenecks limit our understanding of point sources and their role in the global carbon cycle. This thesis presents FootNet, a machine learning surrogate model designed to emulate atmospheric transport at a fraction of computational cost with “on-the-fly” computation capabilities. The performance of FootNet is first demonstrated in an urban case study using dense in-situ CO2 observations from the San Francisco Bay Area. Results show that ML-based inversions not only reproduce spatial emission patterns consistent with physics-based models but also improve agreement with independent observations, likely due to reduced sensitivity to transport errors. FootNet is then generalized to simulate out-of-sample atmospheric transport across the Contiguous United States, maintaining strong performance for both surface and satellite measurements. This generalized version of FootNet matches or exceeds performance in out-of-sample GHG flux inversions for case studies using surface-level and satellite observations. Feature importance analysis indicates that the model learns physically meaningful drivers, such as wind fields and the Gaussian plume, enabling robust generalization. Finally, the ML-based inversion framework is applied to quantify methane emissions in the Permian Basin using TROPOMI methane observations. The high-resolution inversions reveal strong spatial correspondence between inferred emissions and oil & gas infrastructure, including wells, pipelines, compressor stations, storage tanks, and processing facilities. We find total Permian Basin emissions of 2.6 ± 0.5 Tg/yr, a factor of 2.4 higher than bottom-up estimates. We observe a decline in methane intensity from 5% in 2018 to 2% in 2024, while the emissions remained stable and production scaled by 2-3x over the past seven years. Overall, this work demonstrates that machine learning-based emulators of atmospheric transport can overcome longstanding computational barriers, enabling near-real-time, high-resolution monitoring of greenhouse gas emissions. This capability provides a scalable pathway for identifying emission hotspots and supporting data-driven climate policy and mitigation efforts.