The authors use a data-driven approach (based on long-term satellite observations + neural networks + Monte Carlo methods) to study the impact of Arctic sea ice cover on absorbing aerosol direct radiative forcing (ADRF) and reveal its long-term trend. The following issues should be addressed before publication:

1. The study’s methodology is not highly innovative, as NN + data has been used in previous studies. However, the scientific findings seem more novel, particularly regarding the impact of sea ice on ADRF and the long-term trend. The literature review should more clearly compare this work with existing studies to confirm the novelty of the sea ice-ADRF relationship, as many GCM studies have already estimated Arctic ADRF and explored aerosol-sea ice interactions.
2. The 50+ samples used to train the NN might not be sufficient to represent all atmospheric conditions in the Arctic fully. If the NN primarily learns radiation fluxes from low-aerosol regions while the studied region has aerosols at higher altitudes, SWFcln may have systematic bias, leading to ADRF estimation errors. Stronger independent validation is needed to ensure the reliability of NN predictions.
3. In the section on neural network design on pages 11-12 (lines 300-315), the authors mention: “All nodes in the hidden layer use the Leaky Rectified Linear Unit (LeakyReLU) activation (Maas et al., 2013), with this activation function having been identified to provide the best performance after testing with other activation functions.” Why does LeakyReLU provide the best performance？ It is suggested that comparative test results for different activation functions (such as ReLU, Sigmoid, etc.) be provided and that the specific advantages of LeakyReLU in handling TOA radiation flux estimation be explained.
4. Monte Carlo methods quantify uncertainty but cannot verify potential systematic biases. How can the authors confirm that NN-predicted SWFcln has no systematic bias?
5. Can Figure 2 include quantitative data on misclassification, such as the percentage of smoke misidentified as clouds?
6. Why does ADRF shift from negative (cooling) to positive (warming) at a critical sea ice concentration of approximately 60%? Why is 60% the turning point? How does aerosol-surface multiple scattering influence ADRF?

This study presents an interesting data-driven approach, combining satellite observations, neural networks, and Monte Carlo uncertainty estimation, to derive Arctic absorbing aerosol direct radiative forcing (ADRF) trends over a 15-year period. The topic is highly relevant and timely, given the sensitivity of the Arctic climate. The manuscript is generally clear, but several areas require improvement to strengthen its scientific rigor and clarity.

1. Data quality is critical for this analysis. Although the authors utilize several satellite products, many of these have primarily been validated over low- to mid-latitudes. Thorough validation over the Arctic region is necessary. More importantly, uncertainties associated with cloud, aerosol, and surface classification must be quantified. How reliable are the cloud-free and aerosol-free conditions as defined? Similar validation is needed for other aerosol and cloud products.
2. The authors should clarify whether all retrieved data were used or if any quality control measures (e.g., quality flags) were applied.
3. The dataset used for training the neural network appears limited, which could introduce biases, particularly given the uncertain data quality. This needs careful discussion.
4. Figure 5 is not very informative for understanding the neural network architecture. A clearer schematic illustrating the network structure and flow is recommended.
5. CERES data are used as the reference for shortwave flux (SWF) validation. However, the authors must first assess and validate the accuracy of CERES data specifically over the Arctic.
6. The method used for trend estimation should be described in detail. Was data uncertainty incorporated into the trend analysis? In Section 4.3, the error analysis is not pixel-based—how representative is this approach?
7. On line 564, the authors assume that daily Level 3 ADRF errors are normally distributed. This assumption should be justified with supporting analysis.