FCNNs trained on three SN Ia models and evaluated on all samples of its corresponding test set, in comparison to the evaluation on all samples from four test sets from the four individual SN Ia models. The left column shows the case where the SN Ia models have been used as shown in Fig. 1. The right column contains plots in which the SN Ia models are randomly shifted within ±5 d in time and ±0.4 in magnitude. The random shifts make the training set generalize better to test sets that are not represented by the training set.