Skip to main content
Back to timeline
arXivSource publication:

AuroraChem puts chemistry and weather in one differentiable model: a one-year rollout takes 12 minutes, and weather-only insertion lifts OH correlation from 0.89 to 0.93

Synopsis

Joint chemical and meteorological data assimilation has been limited by the cost of coupled chemistry-climate models, which must supply chemical and meteorological states together with their covariances. AuroraChem fine-tunes the Aurora atmospheric foundation model on GEOS-Chem chemical fields and ERA5 meteorology, running a one-year rollout in about 12 minutes on one GPU; free rollouts stay stable for a year but do not fully capture seasonal cycles. In idealized direct-insertion tests, inserting only meteorological fields raised the annual-mean spatial correlation of 500 hPa OH from 0.89 to 0.93, close to 0.94 when inserting the chemical fields themselves, while inserting only chemical fields raised 500 hPa temperature from 0.89 to 0.92.

Source-provided article image: Enabling joint chemical and meteorological data assimilation with an atmospheric foundation model
arXiv

Interpretation

A fine-tuned atmospheric foundation model can supply meteorological and chemical states together, making global, multi-season joint chemical and meteorological data assimilation computationally practical. Coupled chemistry-climate models provide such a joint state, but adding interactive chemistry to CESM2 raises its cost sevenfold, and ensemble assimilation multiplies that by 10 to 100 members, putting global multi-year assimilation out of reach. A one-year AuroraChem rollout takes about 12 minutes on one H100 GPU, against about 4,000 minutes for a one-year GEOS-Chem simulation on a 32-core CPU, though GEOS-Chem simulates roughly 300 chemical species on 72 levels while AuroraChem carries 7 chemical and 5 meteorological fields on 13 levels.

Chemical and meteorological observations can each constrain the other's state in AuroraChem, which is the cross-variable information joint assimilation needs. Earlier chemical assimilation mostly used chemical transport models driven by offline meteorology, so the chemical state did not feed back to the climate state, and regional studies covered only days. In out-of-sample 2019 direct-insertion tests, inserting only chemical fields kept 500 hPa temperature at an annual mean correlation of 0.92 versus 0.89 for the free rollout and recovered the Indian summer monsoon moisture and Somali jet; inserting only meteorological fields kept 500 hPa OH at 0.93 versus 0.89 free and 0.94 when inserting chemical fields, and recovered the OH seasonal cycle.

Sparse column observations of a single species can also constrain OH, opening a path for real satellite observations into assimilation. Earlier satellite-based OH estimates were limited by the sparsity of retrievals, covering only open or tropical oceans at monthly resolution. Inserting only NO2 gave an annual-mean 500 hPa OH correlation of 0.95; sampling at TROPOMI pixels and inserting only columns computed with a TROPOMI averaging kernel each gave 0.93, against 0.89 for the free rollout.

Perspective

The tests are idealized direct insertion on out-of-sample 2019, with pseudo-observations taken from truth and no noise; chemical truth comes from a single 2018–2019 GEOS-Chem simulation and meteorological truth from ERA5. The authors propose as the next step assimilating real meteorological and chemical observations, including TROPOMI NO2, to estimate three-dimensional, time-resolved OH.

The authors judge that AuroraChem mainly learns the correlation structure between chemical and meteorological fields rather than the physics coupling them, because chemistry does not drive the monsoon and the training data contain no chemistry-to-meteorology feedback. The model has seen only one year of monthly mean emissions, so it is unsuited to perturbation experiments attributing OH changes to emission drivers, and free rollouts do not fully capture seasonal cycles, with long-lived species developing tropical artifacts.