This paper looks at the sources of ozone in the remote troposphere — the lower part of the atmosphere far from cities and fires. Using a new deep‑learning method, the authors find that fossil fuel combustion supplies more than three times as much ozone there as biomass burning (wildfires and other burning of plant material). This result helps resolve a recent disagreement between aircraft‑based tracer studies and global atmospheric models.

The team first examined why a common observation method (the “OT” tracer approach) suggested that biomass burning made much more ozone than fossil fuels. That method uses chemical tracers such as hydrogen cyanide (HCN) for fires and tetrachloroethylene (C2Cl4) for fossil fuel sources, together with carbon monoxide (CO). The authors show that differences in how long these tracers survive in the air can bias the result. For example, HCN lasts longer than CO, so after long transport the ratio used by the OT method can be artificially large. In one test case over the Atlantic for December 2017–January 2018, the OT method inferred unrealistically long transit times (about 80–180 days) compared with trajectory estimates (2–28 days), which helped explain the inflated fire signal.

To overcome these problems, the researchers built a deep‑learning (DL) framework that learns the link between tracer measurements and source‑specific ozone from a large set of model simulations. They trained the DL model on global chemical transport model (CTM) output (the GEOS‑Chem model) run under several emission scenarios, including a baseline and perturbed cases that changed biomass‑burning emissions by a factor of ten and fossil‑fuel emissions by a factor of two. The DL model was then given observed tracer data from the ATom aircraft campaign to infer how much ozone came from biomass burning versus fossil fuels.

The deep‑learning approach performed well in tests against held‑out model data, with coefficients of determination (R²) above 0.93 and very small bias. By contrast, the OT method showed more scatter and bias (R² about 0.69 for fire contributions and 0.77 for fossil‑fuel contributions, with a large positive bias in the biomass‑burning estimate). Applied to observations, the DL results show a consistently larger influence from fossil fuels than from biomass burning across most regions and seasons.