Benchmarking machine learning models for arctic ozone profile retrieval from MAX-DOAS observations

Continuous, high-resolution ozone (O₃) profiles are crucial for understanding atmospheric chemistry in the climate-sensitive Arctic, yet obtaining such data remains a significant challenge. This study presents a comprehensive benchmark of four machine learning models—Ridge Regression (RR), Random Forest (RF), LightGBM (LGBM), and a Multi-Layer Perceptron (MLP)—for retrieving vertical ozone profiles by fusing ground-based Multi-Axis Differential Optical Absorption Spectroscopy (MAX-DOAS) observations with reanalysis data. The models were trained on a multi-year dataset and evaluated on an independent test period. The performance of each algorithm was assessed using three statistical metrics, including R, RMSE, and MAPE. The results demonstrate the superiority of non-linear approaches over the linear baseline, with LGBM consistently emerging as the superior algorithm and achieving the highest overall accuracy against reanalysis data (R=0.98, RMSE=0.80 mPa). Additionally, its robustness was confirmed through validation against independent measurements, where it exhibited the strongest agreement with both 'gold standard' ozonesondes (R=0.96, RMSE=0.36 mPa) and in-situ surface ozone data (R=0.51, RMSE=6.66 ppbv). By quantitatively evaluating the trade-offs between different modeling approaches, this work provides a clear, evidence-based guide for algorithm selection in ozone retrieval. It establishes that state-of-the-art tree-based ensembles are a computationally efficient and highly accurate tool, offering a reliable pathway for generating the high-time-resolution ozone profile data needed for Arctic research.