With the improved spatial resolution of satellite spectrometers such as TROPOMI, more homogeneous cloudy scenes can be identified at the pixel scale, making scattering-cloud retrievals increasingly worthwhile for satellite cloud processing. DISAMAR provides physically rigorous and spectrally accurate simulations for cloud retrieval from oxygen absorption bands, but its line-by-line radiative transfer calculations are computationally too expensive for efficient large-scale application. In this study, we develop a fast cloud retrieval framework for TROPOMI O<sub>2</sub>-A band observations by combining DISAMAR with neural-network emulators. The method supports the joint retrieval of cloud optical thickness (COT) and cloud optical centroid pressure (CLP) within an optimal-estimation framework. The neural networks are trained offline using a synthetic dataset generated by DISAMAR under representative cloud, surface, meteorological, and viewing conditions. The dataset contains about 400,000 cases, with meteorological inputs from ERA5 and observation geometries sampled from a TROPOMI orbit; about 70% of the cases correspond to land and 30% to water surfaces. The networks emulate both top-of-atmosphere reflectance and its derivatives with respect to the retrieval state variables, thereby replacing the most time-consuming part of the forward model. Initial experiments for fully cloudy O<sub>2</sub>-A cases show that the proposed framework reproduces the main spectral and retrieval behaviour of the DISAMAR-based scheme while reducing the runtime from several hours per case to a few seconds. The proposed framework provides a practical pathway toward fast and physically consistent retrievals of cloud optical properties from TROPOMI O<sub>2</sub>-A observations.