BRONet v1.0: A machine learning approach for BrO profile retrieval using a full-physics inverse model

Abstract. To address the challenge in retrieving tropospheric BrO profiles using the optimal estimation algorithm for Multi-Axis Differential Optical Absorption Spectroscopy (MAX-DOAS)—which relies on a priori profiles—a robust, data-driven retrieval framework, hereafter BRONet v1.0, is developed. First, a comprehensive synthetic dataset comprising 51.5 million samples is generated by coupling physics-based parameterizations of BrO and aerosol profiles with radiative transfer simulations. Second, a hybrid convolutional and long short-term memory neural network (CNN-LSTM) is designed and trained to directly map BrO differential slant column densities (DSCDs) to BrO profiles. This architecture leverages convolutional layers for local feature extraction from multi-angle data and LSTM layers to capture vertical dependencies, enhanced by an attention mechanism for adaptive weighting. The model achieves a coefficient of determination of 0.84 and a root mean square error of 0.36. Finally, the retrieved BrO profiles at Huaniao Island, China, show an average relative deviation of 15.8 % compared with those from the optimal estimation algorithm. Notably, the model completes a single profile retrieval in approximately 0.5 s, representing a 240-fold improvement in computational efficiency over the traditional optimal estimation approach, while delivering an enhanced vertical resolution of 191 layers with 10 m resolution below 1 km within the 0–4 km range.