Explain the poorer recovery of M_cool0 relative to α_LW under SKA-low AA* noise

Determine why the Convolutional Neural Network trained on multi-frequency angular power spectrum (MAPS) maps with SKA-low AA* thermal noise recovers the Lyman–Werner feedback efficiency parameter α_LW with better performance than the baseline H2-cooling threshold parameter M_cool0, and assess whether the redshift evolution of the Lyman–Werner background indeed provides more informative features for α_LW than for M_cool0 in the presence of noise and parameter degeneracies.

Background

The paper trains a Convolutional Neural Network (CNN) on simulated multi-frequency angular power spectrum (MAPS) maps of the 21 cm signal to infer two parameters controlling H2 cooling and Lyman–Werner (LW) feedback: the baseline cooling threshold M_cool0 and the LW feedback efficiency α_LW. Without instrumental noise, both parameters are recovered accurately.

When SKA-low AA* thermal noise is added, the overall recovery degrades and the performance diverges between the two parameters: α_LW maintains moderate predictive power (R^2≈0.54), whereas M_cool0 performs worse (R^2≈0.41). The authors note that the reason for this disparity is unclear, hypothesizing that α_LW may carry more information through its imprint on the redshift evolution of the LW background.

Understanding the cause of this performance gap is important for designing inference strategies and for clarifying which aspects of the VAO-driven MAPS carry robust information about cooling and feedback under realistic noise conditions.