MiRS-AI: A Data-Driven Framework for Accurate and Efficient Atmospheric Profile Retrieval from Microwave Observations

Radiative transfer–based retrieval systems such as the National Oceanic and Atmospheric Administration (NOAA)’s Microwave Integrated Retrieval System (MiRS) are widely used to estimate atmospheric states from satellite observations but are computationally expensive and limited by the ill-posed nature of the inversion problem. We present MiRS-AI, a deep learning framework that formulates atmospheric retrieval as a data-driven surrogate modeling task. The framework includes a pixel-wise ResNet model that emulates traditional 1D-Var retrieval and a spatial UNet-Transformer (UNT) model that captures multiscale dependencies across neighboring observations. Trained on global Advanced Technology Microwave Sounder (ATMS) observations and ECMWF analysis data, MiRS-AI achieves improved accuracy with lower bias and variance than operational MiRS while providing orders-of-magnitude speedup. The UNT model further improves performance in dynamically complex regions such as tropical cyclones, highlighting the importance of spatial context, though it may partially inherit smoothing from reference data and underrepresent fine-scale structures. These results demonstrate that data-driven surrogate models can effectively approximate physics-based retrieval systems, offering a scalable path for future extensions to additional atmospheric parameters and multimodal observations.