Reionisation time field reconstruction from 21-cm maps: Investigating predictor coherence in warm dark matter cosmology

The reionisation time field TR captures the entire history of cosmic reionisation by mapping the moment where each region of the Universe became ionised. Previous work has shown that TR can be inferred from 21-cm observations, including SKA-like instrumental effects, using convolutional neural networks (CNNs). However, these CNN predictors are trained on specific reionisation models, raising critical concerns about their reliability when applied to observational data potentially differing from their training assumptions. This paper proposes and tests a method to evaluate the coherence of our CNN predictors with respect to their input model, thereby enabling the validation or exclusion of underlying reionisation models based on their reconstruction behaviour. We set the cold dark matter (CDM) model as a reference input, and we evaluated the coherence of TR reconstructions by comparing them across different redshifts for several prediction models, as the statistics of TR reconstructions should be the same for every redshift of the input maps. Focusing on metrics such as the root mean square error, the coherence fraction q, and the coefficient of determination R_Δ^2, our study specifically investigates CNNs trained on CDM and warm dark matter (WDM) models, with WDM particle masses of 2, 3, 5, and 7 keV. We find that the predictors trained on 5 and 7 keV WDM models exhibit high-level self-consistency similar to the CDM predictor, while the 2 keV predictor and (to a lesser extent) the 3 keV predictor display significant deviations across several metrics. These findings seem to demonstrate that CNN predictors retain sensitivity to differences in the underlying reionisation model and can be used to assess model compatibility with observations. Our results highlight the necessity of validating machine-learning predictors against their input models before applying them to real data. The method proposed here offers a pathway to more reliable applications of CNNs in the study of reionisation, and future work will be aimed at enhancing the model's performance through improvements in CNN architectures and the adaptation to SKA-like observations.