The subject matter of the present invention relates to trained machine-learning models (300), methods (200, 400) and apparatuses (500) allowing a future resonant frequency of a catalyst for selective reduction of nitrogen oxides (SCR) to be predicted, the resonant frequency being representative of a concentration of a reducing agent within the SCR. The SCR forms part of a system for after-treatment of a flow of exhaust gases of an internal combustion engine with which a motor vehicle is provided. The general principle of the invention is based on the observation of correlations between the resonant frequency of an SCR and the concentration of ammonia present within the SCR. This observation led the inventor to envision using machine learning to create a trained machine-learning model in order to predict the resonant frequency of an SCR. In the invention, the trained machine-learning model is a so-called predictive model in which significant correlations are discovered in a set of past observations and in which it is sought to generalize these correlations to cases that have not yet been observed.