Expositions of Sonata Forms are structured towards two cadential goals, one being the Medial Caesura (MC). The MC is a gap in the musical texture between the Transition zone (TR) and the Secondary thematic zone (S). It appears as a climax of energy accumulation initiated by the TR, dividing the Exposition in two parts. We introduce high-level features relevant to formalize this energy gain and to identify MCs. These features concern rhythmic, harmonic and textural aspects of the music and characterize either the MC, its preparation or the texture contrast between TR and S. They are used to train a LSTM neural network on a corpus of 27 movements of string quartets written by Mozart. The model correctly locates the MCs on 14 movements within a leave-one-piece-out validation strategy. We discuss these results and how the network manages to model such structural breaks.