Context. In the context of structure formation, disentangling the central galaxy stellar population from the stellar intrahalo light can help us shed light on the formation history of the halo as a whole, as the properties of the stellar components are expected to retain traces of the formation history. Many approaches are adopted to assess the task, depending on different physical assumptions (e.g. the light profile, chemical composition, and kinematical differences) and depending on whether the full six-dimensional phase-space information is known (much like in simulations) or whether one analyses projected quantities (i.e. observations). Aims. This paper paves the way for a new approach to bridge the gap between observational and simulation methods. We propose the use of projected kinematical information from stars in simulations in combination with deep learning to create a robust method for identifying intrahalo light in observational data to enhance understanding and consistency in studying the process of galaxy formation. Methods. Using deep learning techniques, particularly a convolutional neural network called U-Net, we developed a methodology for predicting these contributions in simulated galaxy cluster images. We created a sample of mock images from hydrodynamical simulations (including masking of the interlopers) to train, validate and test the network. Reinforced training (Attention U-Net) was used to improve the first results, as the innermost central regions of the mock images consistently overestimate the stellar intrahalo contribution. Results. Our work shows that adequate training over a representative sample of mock images can lead to good predictions of the intrahalo light distribution. The model is mildly dependent on the training size and its predictions are less accurate when applied to mock images from different simulations. However, the main features (spatial scales and gradients of the stellar fractions) are recovered for all tests. While the method presented here should be considered as a proof of concept, future work (e.g. generating more realistic mock observations) is required to enable the application of the proposed model to observational data.