Diagnostics of the performance of the emulator. Top left: At each k, z, median across the dataset of the emulator error-to-thermal-noise ratio. Top right: At each k, z, median across the dataset of the emulator error-to-cosmic-variance ratio. The error increases with cosmic variance, as the deterministic network cannot replicate the noise induced by cosmic variance. Bottom left: 100x mean squared relative error across the dataset between the emulator prediction and the training sample. Bottom right: Loss at each k, z. Faint signal at high z causes low loss but high relative error. Cosmic variance causes high loss and relative error at low k. Additionally, we checked that the mean square relative error exhibits no systematic trend depending on the values of the astrophysical parameters.