Abstract. The relationship between geostationary radiances and ice water path (IWP) is complex, and traditional retrieval approaches are not optimal. This work applies machine learning to improve the IWP retrieval from Meteosat-9 observations, with a focus on low latitudes, training the models against retrievals based on CloudSat. Advantages of machine learning include avoiding explicit physical assumptions on the data, an efficient use of information from all channels, and easily leveraging spatial information. Thermal infrared (IR) retrievals are used as input to achieve a performance independent of the solar angle. They are compared with retrievals including solar reflectances as well as a subset of IR channels for compatibility with historical sensors. The retrievals are accomplished with quantile regression neural networks. This network type provides case-specific uncertainty estimates, compatible with non-Gaussian errors, and is flexible enough to be applied to different network architectures. Spatial information is incorporated into the network through a convolutional neural network (CNN) architecture. This choice outperforms architectures that only work pixelwise. In fact, the CNN shows a good retrieval performance by using only IR channels. This makes it possible to compute diurnal cycles, a problem that CloudSat cannot resolve due to its limited temporal and spatial sampling. These retrievals compare favourably with IWP retrievals in CLAAS, a dataset based on a traditional approach. These results highlight the possibilities to overcome limitations from physics-based approaches using machine learning while providing efficient, probabilistic IWP retrieval methods. Moreover, they suggest this first work can be extended to higher latitudes as well as that geostationary data can be considered as a complement to the upcoming Ice Cloud Imager mission, for example, to bridge the gap in temporal sampling with respect to space-based radars.