Review of "Interpretable Machine Learning Quantifies Composition and Size Controls on Aerosol Spectral Absorption"

This manuscript by Wang et al. investigates the complex drivers of the Absorption Ångström Exponent (AAE) using a combination of high-resolution field observations in Beijing and multi-year AERONET columnar data. The study addresses a persistent challenge in atmospheric science: quantitatively disentangling the relative importance of chemical composition versus particle size in determining aerosol spectral absorption. The authors employ Shapley Additive exPlanations (SHAP) to provide a ranked attribution of AAE drivers. The study quantifies how AAE variations directly influence aerosol radiative forcing efficiency (ARFE), finding that AAE is a primary driver of cooling efficiency at the Top of the Atmosphere (TOA). The manuscript is well-structured and the methodology is fairly robust, with the caveats detailed below.

Major issues

1. The study relies heavily on SHAP to attribute drivers of AAE which identifies associations, not necessarily causal links. The authors acknowledge nonlinearity and collinearity among predictors, however, if two variables are highly correlated (e.g., $D_{APS}$ and FMD mass fraction, $r=0.64$), SHAP can sometimes "split" the importance between them in ways that don't reflect physical reality. The manuscript would benefit from a more explicit discussion on whether the "importance" found by the model is supported by Mie theory or other physical optical models to bridge the gap between statistical importance and physical causation.
2. Near-surface chemical data was collected via offline filter sampling (day/night blocks), while optical data was online (hourly). While the authors matched these temporally, the coarse resolution of the chemical data (12-hour averages) likely masks the fine-scale diurnal variability seen in the optical $AAE_{sfc}$.
3. The authors describe using a ‘consistent split’ for model training, testing and validation but never explicitly state what that is. Please describe the method for splitting and why that’s appropriate for this dataset.

Minor issues

1. The authors focus on a specific site in Beijing, which given the observational constraints seems reasonable, but it would be beneficial to include a sentence or two in the conclusion explicitly stating how these results might change (and therefore their relevance) in cleaner or more dust-dominant global regions.
2. Please mention the ranges quoted in the results section. Are these 1 standard deviation, or something else?
3. L217, "coare-mode" should be corrected to "coarse-mode".

This manuscript by Wang et al. investigates the relative importance of aerosol chemical composition and particle size in determining the Absorption Ångström Exponent (AAE), using both ground-based observations and column-integrated AERONET data within an interpretable machine learning framework (SHAP). This study presents a novel and insightful application of machine learning to a challenging problem in aerosol science. The quantification of the relative roles of composition and size on AAE is a significant contribution. However, the manuscript requires major revisions to address the critical issue of causality in the radiative forcing analysis, clarify key methodological steps (especially the temporal matching), and provide a more integrated and critical discussion of the multi-platform datasets. With these revisions, the paper will be well-suited for publication in Atmospheric Chemistry and Physics.

Major Comments

1. Causality and the Role of AAE in Radiative Forcing (Lines 224-242, 424-459, and Section 4):
   
   The most significant conceptual issue lies in the framing of AAE's role in radiative forcing. The authors state that they aim to "elucidate the critical role of AAE in radiative effects" and later use SHAP to quantify AAE as a "driver" of ADRF and ARFE variations. This implies a causal relationship where AAE is an independent variable controlling radiative forcing.However, as the authors themselves expertly demonstrate in the first half of the paper, AAE is not a fundamental physical property; it is a diagnostic metric that is itself driven by the same factors that control radiative forcing: composition (BC, BrC, dust) and size. To then turn around and treat AAE as an independent "driver" of radiative effects is circular. The fundamental drivers are the microphysical and chemical properties (e.g., BC concentration, dust loading, fine-mode radius). AAE is a valuable observational constraint precisely because it integrates these properties, but it is a consequence, not a cause. The analysis of AAE's relationship with DRF is still scientifically valuable, but its framing must be corrected. The study should not claim to quantify AAE's "driving" role. Instead, it should be framed as an investigation into the diagnostic power of AAE. The goal should be to understand how well this convenient, measurable optical parameter can explain or predict variability in radiative forcing. For example, showing that AAE is a strong predictor of TOA forcing efficiency (as in Figure 7) is a useful result: it suggests that if you can measure or constrain AAE, you have a powerful tool for estimating the aerosol's radiative effect. This is a correlative/ diagnostic relationship, not a causal one. The language throughout Sections 3.4 and 4 must be revised to reflect this, replacing terms like "driver," "regulator," and "governs" with phrases like "is associated with," "can help predict," or "serves as a key diagnostic for."
2. Matching Offline and Online Measurements for Model Training (Section 2.2, Lines 175-186):
   
   A critical methodological detail is insufficiently explained. The multiple linear regression model in Section 2.2 uses offline chemical composition data (FMD fraction, nd-WSII fraction) matched with online AAE and size distribution data. The authors state that data were "temporally matched to the corresponding online measurements based on sampling periods" (daytime 09:00-20:30; nighttime 21:00-08:30).This averaging over ~11.5-hour and ~11.5-hour periods is a significant source of uncertainty and potential bias. Within a single daytime or nighttime filter sample, the aerosol composition, size distribution, and AAE are likely highly variable due to changes in emissions (e.g., rush hour), boundary layer dynamics, and chemistry. Assigning a single, averaged composition value to the highly temporally resolved online data within that period assumes a static relationship that may not hold. The authors must: （1）Explicitly justify why this temporal resolution is sufficient to capture the relationships they are investigating. （2）Discuss the potential for "ecological fallacy" or averaging bias—where the relationship between variables at an aggregated level differs from the relationship at a high-resolution level. （3）Ideally, provide an uncertainty estimate for how this temporal mismatch might affect the regression coefficients and conclusions.
3. Disconnect Between Surface and Columnar Models (Sections 3.2 and 3.3):
   
   The manuscript essentially presents two independent modeling efforts: one for surface AAE (using MLR with 4 predictors) and one for columnar AAE (using ML models with 9 predictors). The connection between these two parts is weak. The surface analysis uses direct physical measurements, while the columnar analysis uses retrieved optical properties and a chemically inverted dataset. The manuscript would be strengthened by a more explicit discussion of how these two perspectives complement each other. For instance, does the dominant role of dust at the surface (FMD) align with the importance of CAI (coarse-mode absorbing dust) in the column? How do the limitations of one dataset inform the interpretation of the other? A dedicated paragraph synthesizing these findings and acknowledging their different physical meanings would greatly improve the manuscript's coherence.
4. Uncertainty in AERONET-Inverted Chemical Composition (Section 2.3):
   
   The analysis of columnar AAE relies heavily on the AERONET chemical composition product (BC, BrC, CAI, etc.). It is crucial to remind readers that these are not directly measured but are retrieved from inversions of spectral sun photometer measurements, which come with their own assumptions and uncertainties. The manuscript briefly cites Zhang et al. (2024), but a more critical discussion is warranted here, especially given the central role of these data in Figure 5. What are the primary assumptions in this retrieval? (e.g., regarding refractive indices, mixing state, particle shape). What is the estimated uncertainty for each component (BrC, BC, dust) as provided by the retrieval algorithm or the literature? A short statement acknowledging these limitations and citing key references on the uncertainties of AERONET inversions (e.g., Dubovik et al., 2000; Sinyuk et al., 2020) would provide necessary context for the robustness of the SHAP results.

Minor Comments

1. Line 89-90: The phrasing "an ensemble of models was initially trained, after which the optimal model was selected" is slightly ambiguous. Clarify that you trained multiple model types and selected the best-performing one (CatBoost) for the final interpretation, as described later in Section 2.4. This is good practice, but the wording could be more precise.
2. Line 245-247: The acronyms SMPS and APS are used but were introduced in Section 2.1.1. Since this is the start of the Results section, it might be helpful to briefly re-introduce them as "fine-mode (SMPS) and coarse-mode (APS) particle sizers" for readers who may not have the methods section fresh in mind.
3. Line 340-341: The sentence "The AAEcol (1.47±0.56) was also suggested to be greater than that derived from the surface field campaign" is a bit awkward. Replacing "was also suggested to be" with "was found to be" or "was also higher than" would be clearer.
4. Line 384: "explaining ∼50%of model performance." It would be more precise to say "explaining ∼50% of the model's predictive power (as measured by mean absolute SHAP value)" or something similar, as SHAP importance sums to the total model output, not necessarily a performance metric like R².
5. Figures 6 and 7: The box plots in (a-c) are very effective. The SHAP summaries in (d-f) are informative. Consider adding the sample size (n) for each AAE bin in the box plots to give the reader a sense of the statistical robustness of each category.
6. Line 225: Change "in determining" to "in predicting" or "in explaining the model's estimation of."
7. Line 424: Change "revealed a robust, layer-dependent coupling" to "revealed a robust, layer-dependent correlation."
8. Line 477: Change "is a key regulator of" to "is a strong predictor of" or "contains valuable information for estimating."
9. Line 484-486: Change "demonstrate the multifactorial control of AAE" to "demonstrate the multifactorial influences on AAE" and "highlight its pivotal role in partitioning radiative forcing" to "highlight its strong correlation with the vertical partitioning of radiative forcing."