Left: total variance recovered as function of number of principle components. We show Lyα line profiles spamming zELDA’s grid without (with) IGM absorption in red (green). The dashed black line marks the number of principal components used for the input of the artificial neural networks. We fixed the number of PCA to 100 to achieve a re-coverage of 95% of the total variance. Right: example of PCA decomposition in an IGM clean-shell model line profile. The line profiles are shown in the proxy rest frame. The original mock line profile is shown in grey. Meanwhile, the reconstructed line profiles using the N first principal components are displayed in the legend.