The current advancements of noninvasive imaging methods applied for the study and conservation ofcultural heritage have driven a rapid development of novel computational methods. Macro X-rayfluorescence (MA-XRF) is well-established and used for the investigation of paintings. However, MA-XRFgenerates large datasets that can be challenging to analyze. In the following, we employ machinelearning approaches for the analysis as they allow for identification of non-trivial dependencies andclassification across the high dimensional data, hence promising comprehensive interrogation. We havebuilt a novel deep learning algorithm trained on a synthetic dataset that allows for fast and accurateanalysis of the XRF spectra in MA-XRF datasets circumventing typical drawbacks of the classicaldeconvolutional approach. The synthetic XRF spectra are generated using Monte Carlo simulations basedon a Fundamental Parameters approach and tuned for our MA-XRF setup. The simulations assume astratigraphy model of a painting with a large number of possible historical and modern pigments. Thepresented approach yields high-quality results in terms of analysis of MA-XRF scans.Here we focus on possible extensions of the methodology and its advanced applications. First, we extendthe methodology to paintings with detected underpaintings. The neural network is modified to includeadditional parameters relevant to the composition and layering structure. We de-layer the painting andwe separate underpaintings from the visible pictorial composition. Second, we introduce a new AI/MLscheme considering neighboring spectra to increase the detection limit that allows us to improve theimaging quality of the elements with low counts. The examples demonstrate the possibilities goingbeyond conventional approaches. We discuss the results of these methodologies applied to the analysisof both historical and modern paintings.