Amplitude of the 1σ errors as a function of the NDE’s training size, for a fixed training size of the ResNet set at 40 960 XODs. We can see that the NDE already performs well with a few hundred XODs as a training size. This was to be expected as it was intentionally developed to work in this way. Even though working already with smaller training XODs, we conservatively decided to use 10 000 XODs as our final training sample, given that we already simulated 70 000 XODs.