Bin-wise energy resolution for protons (hydrogen) with energies from 20 MeV to 1 GeV. The left panel shows the results if we apply no selection cuts to our test data, the right panel those for events with $N_{\mathrm{sig}}^{\mathrm{yx}}\ge 3$ and $N_{\mathrm{sig}}^{\mathrm{yz}}\ge 3$. The blue histograms give the performance for one neural network trained over the full data set of (stopping and penetrating) hydrogen nuclei; the yellow and red histograms that of separate networks for stopping and penetrating particles. The gray histograms show the performance of a branched framework, where a filter network separates stopping from penetrating particles and passes them to the separate energy-reconstruction networks.