Interpreting 2D-NMR spectra using Grad-CAM

It has been shown that it is possible to train a (simple) neural network to classify nuclear magnetic resonance spectra by a substructures either being part of the chemical structure measured or not. We now explore the interpretability of such models using techniques from explainable AI, specifically Grad-CAM. We show that those techniques do not give ideal results in the context of NMR, which would be able to identify individual peaks. On the other hand, they enable a better interpretation of the results than those metrics just based on "right or wrong". We can also confirm the result from our previous work, that the trained network performs well for pure compounds, but its generalisability to mixtures is questionable, a limitation that could only be assumed in the original study.