Machine learning-based lidar simulations and retrievals under a pristine clear sky

This study presents a new neural network approach to efficiently simulate lidar signals and retrieve atmospheric profiles. It targets at multi-wavelength, multi-channel lidar systems including backscatter lidar at 355 nm, 532 nm and 1064 nm, as well as Differential Absorption Lidar (DIAL) at 1571 nm for CO2 measurements. A Transformer model is adopted and trained to capture complex, non-linear relationships among those lidar signals and atmospheric properties, especially during backward retrievals, which are not fully describable by traditional physics-based methods. Results show that our approach significantly accelerates computational speed while improving both signal simulations and retrievals accuracy. Specifically, our method uncovers hidden inter-channel physical relationships achieving forward-simulation errors below 1%. When using lidar signals alone, the retrieval uncertainties of CO2 mixing ratio, air density and temperature are below 0.2%, 0.2% and 1K, respectively. It also reduces the simulation time for one satellite orbit from approximately 1 minute (with traditional methods) to only 0.30 seconds.