Inferred model parameters versus input model parameters for case in which spectra have from 1000–10 000 counts spread over 200 bins. The neural network is trained with 20 000 simulations and the posteriors for 500 test spectra are then computed. The medians of the posteriors are computed from 20 000 samples, and the error on the median is computed from the 68% quantile of the distribution. The linear regression coefficient is computed for each parameter over the 500 test samples.