In this paper, the authors applied a sampling-based optimization algorithm first developed by Shi and Sun (Shi, B. and Sun, G.: An adjoint-free algorithm for conditional nonlinear optimal perturbations (CNOPs) via sampling, Nonlinear Processes

in Geophysics, 30, 263–276, https://doi.org/10.5194/npg-30-263-2023, 2023) to ENSO studies. More specifically, the authors applied the algorithms to solve the optimal precursors for the Zebiak-Cane model. Numerical results are also provided in comparison with the traditional adjoint method. 

This is an interesting paper as it demonstrates the practical application of the algorithm proposed by one of the authors to real atmospheric dynamic models. On the whole, I found the exposition and goal of the paper to be clear. The numerical results seem to perform well and the paper is well-written. Therefore I endorse its publication. That said, certain parts of the paper needs minor revision. 

Since it is purely application-based paper and parallel computation is not unique for Monte Carlo methods, I will leave the decisions to the associate editor whether the paper aligns with the scope of Nonlinear Processes in Geophysics.

Comments:

1. In equation (1), I found it more intuitive to keep the unit in the dinominantor if the goal is to nondimensionalize the variable.

2. Table 1: to make it more clear, the authors need to explicitly say that the values reported in table are the original objective values formulated in (4) instead of the sample average (5).

3. Table 2: more discussions are needed for the settings of the experiment. For example how many cores are used in parallel sampling? Is multithreading allowed for the adjoint method? 

4. The original paper Shi and Sun 2023 applied the algorithm to some traditional dynamic system toy models such as viscous Burger's equation and Lorenz 96 model, and it only requires 5-15 samples for the method to perform well in these low-dimensional settings. It would be great if the authors could include some discussions about convergence results and how the number of samples scales with the dimensionality of the models.

Typos:

line 37: "is" -> "was"

line 67: "reivew" -> "review"

page 3 footnote: "constrained optimization"

line 94: "furthre researches" -> "further research"

line 252: "probably" -> "probable"