Correcting aerosol extinction coefficient vertical structure biases in GEOS-chem via a physics-informed transformer with physical mechanism diagnosis

Abstract

Accurately characterizing aerosol vertical distributions is essential for evaluating radiative forcing and air quality. While Chemical Transport Models (CTMs) simulate spatially continuous Aerosol Extinction Coefficient (AEC, km−1), they exhibit systematic AEC biases. Cloud-Aerosol Lidar with Orthogonal Polarization (CALIOP) observations provide precise AEC profiles but are constrained by sparse spatial sampling. To bridge this gap, we propose a physics-informed Transformer framework as a supervised bias-correction model to correct biases in the AEC profiles simulated by GEOS-Chem. Unlike a standard Transformer, our framework features a dual-stream architecture with explicit physical constraints. It employs gated feature fusion to integrate vertical structures (combining GEOS-Chem priors with MERRA-2 profiles) by dynamically identifying height-dependent drivers, and leverages cross-attention to incorporate MERRA-2 surface environmental constraints for modulating AEC vertical rectification with synoptic contexts. This approach effectively predicts systematic biases relative to CALIOP satellite observations and resolves AEC profiles, surpassing methods retrieving only aerosol layer heights. Leave-One-Year-Out validation over East Asia during 2017–2019 demonstrates significant AEC precision improvements, increasing R from 0.49–0.53 in the GEOS-Chem simulations to 0.66–0.73 and reducing RMSE by approximately 25 %. The model effectively mitigates over-diffusion, significantly reducing AEC simulation biases in the critical near-surface layer while capturing smoothed biomass burning and dust plumes. Additionally, it exhibits robust cross-continental transferability, reproducing bias patterns over the North American domain (R=0.70) without retraining, confirming the internalization of universal physicochemical relationships linking atmospheric states to simulation biases. Furthermore, interpretability analysis serves as a diagnostic tool to guide physical model improvement. The model identifies temperature and sensible heat flux as primary drivers to constrain boundary layer mixing, pointing to potential uncertainties in vertical eddy diffusion. Additionally, it uses environmental proxies (e.g., vegetation indices and diffuse radiation) to diagnose potential deficiencies in dust threshold friction velocity and secondary organic aerosol yields. These insights provide a physical basis for refining parameterization schemes in CTMs.