A Physics-Informed Auto-Learning Framework for Developing Stochastic Conceptual Models for ENSO Diversity

The paper explicitly defines and parameterizes the 4D IaIsM (Eq. 8.4, Table 6) with an intraseasonal WWB variable τ driven by state‑dependent multiplicative noise στ(TC), and validates it under a unified 20,000‑year/271×70‑year protocol; it shows that all models except the 3D IaM meet the validation metrics, while the 3D model fails on higher‑order moments, ACFs, and extreme/multi‑year event frequencies. The candidate solution mirrors these results and reasoning. Minor issues: it reverses the reported TC/TE recovery correlations (~0.75 for TC, ~0.82 for TE in the paper), and slightly overstates minimality by implying a universal impossibility for any 3D model within the library, which the paper does not formally prove. Key assumptions in the paper (e.g., adopting the multiplicative noise functional form from the reference model rather than learning it) should be noted.