This manuscript introduces a deep learning approach to estimate the optical properties of internally mixed aerosol particles, considering the non-sphericity of insoluble particles. The research has significant implications for atmospheric models.

Comments for Improvement:

1. Generalizability: While the discussed machine learning (ML) approach demonstrates impressive accuracy, concerns remain about its generalizability. ML-based algorithms often struggle with generalization beyond their training data (Kumar et al., 2024). The authors should address this issue explicitly.

2. Comparison with Observed AOD: Including a comparison of results with observed Aerosol Optical Depth (AOD) would enhance the manuscript. This validation step provides valuable context.

3. Time Complexity: The authors do not discuss the time complexity of the developed approach. One of the main reasons scientists are increasingly pursuing ML is due to its computational cost advantages. 

4. Figure 3: Clarify whether the results in Figure 3 pertain to the test or training dataset. Additionally, provide details on pre-processing and post-processing steps, which are currently missing.

5. Tuning of Hyperparameters: While the network's results are extraordinary, the manuscript lacks information on hyperparameters and their tuning. Including this would improve transparency.

6. Methodology Reorganization: Consider reorganizing the methodology section. Present results from AI-NAOS (lines 174-190, including Figure 2) after describing the approach in detail.

7. Move lines 208-216 and Figure 3 to the results section for better flow.

Reference:

Kumar, P., Vogel, H., Bruckert, J. et al. MieAI: a neural network for calculating optical properties of internally mixed aerosol in atmospheric models. npj Clim Atmos Sci 7, 110 (2024). https://doi.org/10.1038/s41612-024-00652-y

The manuscript introduces AI-NAOS, an AI-based module that incorporates nonspherical and inhomogeneous aerosol particles into radiative transfer simulations. The use of deep learning and advanced optical modeling is an innovative approach in this field, significantly improving accuracy in predicting the direct radiative effects of aerosols. Real-case simulations provide substantial evidence of the AI-NAOS module's impact on atmospheric thermodynamic structures and precipitation patterns, enhancing the understanding of aerosols in weather modeling. In general, the manuscript is well-organized. Only minor revision is needed before publication:

1. The model ran for 72 hours starting from January 12, 2018, with only 24 hours of spin-up time, which seems rather short. Is it possible to extend the model's spin-up time and total run time?
2. Is 'deep neural network (DNN)' used in the text as a general term for a class of methods, or does it refer to a specific neural network method? Please clarify this further in the text.
3. Why were different methods chosen for shortwave radiation and longwave radiation? I believe that RRTM can also be used for shortwave calculations. What considerations influenced this decision?
4. Could you compare it with the latest research on non-spherical black carbon, such as the study by Chen, G., Liu, C., Wang, J., Yin, Y., & Wang, Y., JGRA, 2024, which accounts for the mixing state, nonsphericity, and heterogeneity effects of black carbon in its optical property parameterization within a climate model?
5. I believe there is still considerable uncertainty regarding the impact on precipitation in this section. It's worth further consideration whether to include this in the article. Additionally, while not absolutely necessary, it might be beneficial to examine changes in the boundary layer as well.

Some minor comments:

1. In Line 40, omega is used to represent SSA, while in Line 271, <SSA> is used. Please ensure consistency. The same applies to the asymmetry factor.
2. Line 80: It is usually referred to as 'WRF-Chem' rather than 'WRF/chem'.
3. The vertical axis of Figure 6b is labeled 'atmo. abs.', which seems less rigorous than 'DRE'.
4. Section 3.4 appears twice; the precipitation part should be labeled as 3.5.