DACNO₂ is developed to provide daily 3D NO₂ fields at high spatial resolution (2 km × 2 km) with improved accuracy and generalizability by integrating multi-source data, physically consistent process-based datasets, and real-world measurements. The overall framework, illustrated in Fig. 1, combines diverse data streams with a phased training strategy.

DACNO₂ uses five groups of input features: temporal indicators, emission inventories and proxies, geographic data, ERA5 single-level meteorological variables, and ERA5 multi-level meteorological variables. Together, they provide complementary information on spatial and temporal NO₂ variability. For model training, the targets are process-based NO₂ fields from the CAMS European air quality reanalyses (Inness et al., 2019; Peuch et al., 2022) and real-world surface NO₂ measurements from the EEA AirBase network (European Environment Agency, 2024). CAMS supplies physically consistent large-scale 3D NO₂ distributions, while EEA data constrain the model to match local concentration patterns. Details on data preparation are provided in Sect. 2.2.

To effectively learn NO₂ patterns from diverse datasets, DACNO₂ employs an encoder–decoder architecture with five dedicated encoder branches, each tailored to a specific group of input features. The model structure is described in Sect. 2.3.

Model training is organized into three sequential phases. In Phase 1, a baseline model is pre-trained on process-based CAMS data. In Phase 2, the model is further trained with both process-based and measurement data, improving its ability to represent local NO₂ gradients. In Phase 3, the model is fine-tuned using recent measurements to reflect current NO₂ trends and support real-world applications. Details of the training approach are provided in Sect. 2.4.

The architecture of the DACNO₂ model is illustrated in Fig. 2. The model adopts an encoder–decoder framework with residual connections (He et al., 2016) to map multi-source input features to the daily 3D NO₂ field. The residual connections pass intermediate feature maps from the encoder directly to matching decoder stages, which helps retain fine-scale information across the upscaling path and improves training stability. DACNO₂ integrates several types of neural network modules, including multilayer perceptron (MLP), convolutional neural network (CNN), and convolutional long short-term memory (ConvLSTM), to process and fuse heterogeneous input tensors. Each module is chosen for its specific strengths in handling different data structures. ConvLSTM is for spatiotemporal sequences, CNN is for spatial hierarchies, and MLP is for tabular feature vectors. Inception-style structures are applied in several neural network modules to enable the model to capture both local-scale and broader-scale spatial features.

The DACNO₂ model development employs a three-phase training strategy, including pre-training, multi-constraint fine-tuning, and adaptive fine-tuning. Such a strategy enables the model to learn general patterns (e.g., a-priori knowledge) from a broad dataset and then transfer this internal knowledge to improve its performance on a new, more specific task. Similar approaches have been widely adopted in the development of artificial intelligence (AI) models across various domains, such as earth system modeling, large language models, and biomedical image analysis (Zhuang et al., 2019; Zhou et al., 2017; Ding et al., 2023; Bodnar et al., 2024).

The performance of the DACNO₂ model was evaluated using both EEA NO₂ and CAMS NO₂ test data from 2023. For the comparison against EEA NO₂ (results in the upper panel of Table 2), both DACNO₂ outputs and CAMS NO₂ were evaluated on the 2 km grid. The CAMS NO₂ data was interpolated to 2 km × 2 km resolution (CAMS-2km) and served as the baseline for this comparison. The performance results were calculated across all paired measurements and model estimations. The station-specific time-series consistency analyses are provided in Fig. S3, where the results for each EEA evaluation station were calculated along the daily time-series independently. The average time-series consistency between models and EEA NO₂ is shown in Fig. S4. For the evaluation of DACNO₂ using CAMS NO₂ (results in the lower panel of Table 2), DACNO₂ outputs were evaluated at the CAMS original 10 km × 10 km resolution across all vertical layers, as well as for individual layers. The layer-wise temporal correlations at the regional average and grid scales are illustrated in Fig. S6. Evaluation metrics included the root mean squared error (RMSE), Pearson correlation coefficient (r), coefficient of determination (R2), and bias.

Phases 1–3 represent successive development stages of the DACNO₂ model. The phase-to-phase comparison in Table 2 is used to quantify the incremental effect of adding constraints and the final adaptation step. In Phase-3, the fine-tuning step uses EEA observations from the training stations in 2023. All reported EEA-based metrics are computed on the held-out evaluation stations. Comparisons with EEA NO₂ indicate progressive improvement across DACNO₂ training phases. DACNO₂-Phase-3 achieves the best overall agreement (RMSE = 4.99 µg m⁻³, r=0.82, R2=0.66, bias = -0.38 µg m⁻³), outperforming both DACNO₂-Phase-1 (RMSE = 5.88 µg m⁻³,r=0.75, R2= 0.52, bias = -0.93 µg m⁻³) and DACNO₂-Phase-2 (RMSE = 5.81 µg m⁻³, r=0.79, R2=0.54, bias = 1.67 µg m⁻³). Figure S4 shows that the DACNO₂ model learns reliable temporal correlations with EEA NO₂ at the daily and seasonal scales since Phase-1 (r=0.94), and these correlations are further enhanced in Phase-2 (r=0.95) and Phase-3 (r=0.98). This indicates that the model can represent temporal variability without using satellite NO₂ as an input, relying instead on meteorological and temporal indicators. Moreover, Table 2 and Fig. S4 show a positive bias for DACNO₂-Phase-2 in 2023. This offset is consistent with the fact that the NO₂ level in 2023 is lower than in the Phase-2 training years (2019, 2021, 2022), as illustrated in Fig. S7. DACNO₂-Phase-3 reduces the effect of the interannual variation while maintaining the temporal correlation, highlighting the role of the adaptive fine-tuning step.

Compared to the interpolated CAMS-2km dataset (RMSE = 5.32 µg m⁻³, r=0.80, R2=0.61, bias = -1.15 µg m⁻³), DACNO₂-Phase-3 shows improved accuracy and reduced bias. Station-type analysis further highlights the advantages of DACNO₂-Phase-3, especially at urban and rural sites. For urban locations, DACNO₂-Phase-3 achieves better agreement (RMSE = 5.34 µg m⁻³, r=0.82, R2=0.64, bias = -1.42 µg m⁻³) compared with CAMS-2km (RMSE = 5.89 µg m⁻³, r=0.82, R2=0.56, bias = -2.89 µg m⁻³). In rural areas, DACNO₂-Phase-3 reduces the bias (RMSE = 4.59 µg m⁻³, bias = 1.05 µg m⁻³) compared to CAMS-2km (RMSE = 4.79 µg m⁻³, bias = 1.98 µg m⁻³). Such improvement is consistent with station-specific time-series consistency analysis (Fig. S3). It indicates that DACNO₂-Phase-3 achieves station-specific Pearson correlations comparable to CAMS-2km (CAMS-2km: r-mean = 0.85, r-median = 0.88; DACNO₂-Phase-3: r-mean = 0.84, r-median = 0.87), while exhibiting higher station-specific R2 (CAMS-2km: R2-mean = 0.09, R2-median = 0.52; DACNO₂-Phase-3: R2-mean = 0.23, R2-median = 0.61). The R2 improvements are attributed to more high-R2 sites at urban stations and fewer very low-R2 sites at rural stations. The large difference between the mean and median R2 is primarily caused by negative R2 values at a subset of rural stations, likely due to uneven station distribution across the network and the challenge of modeling weaker, noisier signals in rural environments. Overall, these results suggest that the DACNO₂ model improves agreement with the independent EEA evaluation stations relative to the CAMS-2km baseline across station types. The associated spatial redistribution and localized patterns are examined explicitly in Sect. 3.2.

In addition, Fig. S3 shows that DACNO₂-Phase-3 achieves better station-specific agreement in the central domain than near the boundaries. This may be due to boundary areas that lack sufficient spatial context and have complex mountainous terrain. Additionally, a slight overestimation of DACNO₂-Phase-3 at EEA rural stations persists despite adaptive fine-tuning. A possible reason is the imbalance in the EEA constraint. Figure S1 shows that most stations are located in urban and suburban areas with relatively higher NO₂ concentrations, whereas fewer stations are in rural areas. This may lead to positive bias in the model's estimates for rural areas, and the solutions require further investigation, including sample rebalancing strategies, expanding the study region to include more rural sites, and additional constraints. Meanwhile, given the R2 definition, positive prediction bias at rural stations may be influential, as these stations generally have low NO₂ standard deviations and a smaller tolerance for prediction bias.

Comparisons with CAMS NO₂ across all layers show that DACNO₂ effectively learns and preserves 3D NO₂ distributions through all training phases (Table 2). Figure S6 further presents a layer-wise comparison between DACNO₂-Phase-3 and CAMS at the temporal and grid scales, showing that DACNO₂-Phase-3 can capture the temporal variability of NO₂ in 3D space. Near the surface, DACNO₂-Phase-3 maintains strong agreement with CAMS (Layer 0 m: RMSE = 1.94 µg m⁻³, r = 0.93, R2=0.85, bias = -0.63 µg m⁻³, Table 2), and performance remains robust at mid-altitudes (Layer 500 m: RMSE = 0.79 µg m⁻³, r = 0.90, R2=0.74, bias = 0.05 µg m⁻³), similar to earlier phases. However, a weak correlation is observed in the mountainous region (i.e., the Alps and the Pyrenees, Fig. S6). At higher layers above 1000 m, the agreement in R2 starts to decrease while the correlation remains stable. At 5000 m, DACNO₂-Phase-3 yields a near-zero R2 (-0.01), which is lower than DACNO₂-Phase-1 (0.31) and DACNO₂-Phase-2 (0.34), but the correlation remains moderate (r≥0.6). Figure S6 indicates that the differences between DACNO₂-phase-3 and CAMS at these higher layers are mainly due to magnitude adjustment rather than loss of spatial structure. In addition, predicting very low NO₂ concentrations (approximately 0.05 µg m⁻³ at 3000 m and 0.02 µg m⁻³ at 5000 m, Fig. S2) at high layers is challenged by relatively higher noise. This reduction in agreement at upper layers remains a key challenge for ML-based 3D air quality modeling, which may require additional constraints from space-based observations or physical processes.

To further illustrate the evolution of estimated NO₂ spatial distributions achieved through a phased training, multi-constraint strategy, Fig. 3 compares average surface NO₂ estimates for 2023 from DACNO₂-Phase-1, DACNO₂-Phase-2, DACNO₂-Phase-3, CAMS, and CAMS-2km. Results are shown for the full study region and three representative local areas of Paris, the northern region (NO₂ hotspot area encompassing the Netherlands, Belgium, and the Ruhr area), and the Alpine region.

Across the study region (Fig. 3a), all models exhibit broad and similar NO₂ patterns over land and ocean, consistent with the high spatial agreement between DACNO₂ and CAMS NO₂ reported in Sect. 3.1. Nonetheless, DACNO₂-Phase-2 and DACNO₂-Phase-3 yield sharper spatial contrasts and more clearly defined local NO₂ hotspots than CAMS and DACNO₂-Phase-1. As an additional experiment, we trained the model using only EEA NO₂ data, resulting in the DACNO₂-onlyobs version. As shown in Fig. S8, this model yields effective NO₂ estimates primarily limited to the land surface and cannot reproduce the shipping track patterns, which are visible in the CAMS and DACNO₂ results. Meanwhile, this model produces obvious artifacts over the ocean and at higher altitudes due to the lack of training constraints. These differences highlight the significance of the CAMS NO₂ constraint in facilitating broad spatial generalization in ML-based models.

Differences between models become more pronounced when focusing on local regions (Fig. 3b–d). CAMS NO₂ exhibits visible pixelation effects in these areas due to its coarse native resolution. While bilinear interpolation (as in CAMS-2km) can smooth these effects, it does not introduce additional spatial detail, resulting in oversmoothed patterns. DACNO₂-Phase-1 shows a spatial NO₂ distribution similar to CAMS-2km, despite using high-resolution input features from emission proxies and geography. This indicates that constraints from CAMS NO₂ alone are insufficient for the model to capture fine-scale local NO₂ variability. Incorporating the EEA NO₂ constraint in DACNO₂-Phase-2 addresses this limitation, inspired by approaches in recent ML-based high-resolution surface NO₂ modeling studies using ground measurements as targets (Sun et al., 2024; Wei et al., 2022; Kim et al., 2021; Ghahremanloo et al., 2023). DACNO₂-Phase-2 reconstructs spatial patterns of NO₂ that better match urban layout in Paris (Fig. 3b), identifies more small-scale emission hotspots in the northern region (Fig. 3c), and enhances hotspot signals in the Alpine region (Fig. 3d). DACNO₂-Phase-3 retains these spatial characteristics and primarily adjusts concentration magnitudes by assimilating new measurements to better represent actual NO₂ levels during the application period. For example, the average surface NO₂ concentration estimate in Paris decreases from 11.89 µg m⁻³ in DACNO₂-Phase-2 to 10.08 µg m⁻³ in DACNO₂-Phase-3. This evolution demonstrates the value of integrating multiple constraints and adaptive fine-tuning for high-resolution NO₂ estimation.

To further analyze differences in 3D NO₂ estimates between DACNO₂ and CAMS, Fig. 4 compares their annual average NO₂ distributions for 2023 across all vertical layers over the entire study region and three selected local areas. At the regional scale (Fig. 4a), DACNO₂ and CAMS show strong overall agreement at all altitudes, demonstrating that DACNO₂ effectively learns and reproduces large-scale 3D NO₂ structures from CAMS. However, DACNO₂ provides enhanced spatial detail, presenting sharper gradients and better-defined urban and industrial hotspots, particularly from the surface up to 250 m. At higher altitudes, the differences between the two models gradually diminish, accompanied by a decrease in NO₂ concentrations. Nevertheless, subtle magnitude discrepancies persist, with DACNO₂ estimates reaching lower values, sometimes approaching zero.

Local-scale comparisons further highlight the advantages of DACNO₂ (Fig. 4b–d). In the Paris region (Fig. 4b), DACNO₂ provides finer spatial detail and greater NO₂ levels at lower altitudes (e.g., 0 m: 10.08 µg m⁻³; 50 m: 8.94 µg m⁻³; 250 m: 4.65 µg m⁻³), whereas CAMS results remain coarser with generally lower estimates (0 m: 8.43 µg m⁻³; 50 m: 7.15 µg m⁻³; 250 m: 3.63 µg m⁻³). In the northern region (Fig. 4c), DACNO₂ more distinctly resolves localized emission sources at low layers, capturing a greater number of hotspots than CAMS. As a result, the average NO₂ concentration from DACNO₂ is elevated throughout the boundary layer (up to 1000 m), with mean values 8.8 % higher than those from CAMS. In the Alpine region (Fig. 4d), DACNO₂ more effectively represents terrain-driven gradients and captures NO₂ enhancements within mountainous areas, demonstrating greater sensitivity to complex topographic influences. At higher altitudes, fine-scale variability diminishes in both models and their predicted NO₂ fields become more similar. This is because the influence of local emissions and surface features weakens, while regional-scale processes and long-range transport dominate (see Sect. 4.1). This reduced difference is also accompanied by much lower NO₂ concentrations at higher altitudes.

We assessed the relative importance of input feature groups in DACNO₂ using the integrated gradients (IG) method (Sundararajan et al., 2017) implemented via the Captum interpretability library (Kokhlikyan et al., 2020). IG quantifies the effect of varying each input feature from a zero baseline to its actual value on a selected target function. In this analysis, we computed IG at two targets: (1) the RMSE between DACNO₂ predictions and 2023 EEA NO₂ training measurements at the surface, and (2) the RMSE between DACNO₂ predictions and 2023 CAMS NO₂ at multiple vertical layers. Feature group results are shown in Fig. 5, and results for individual features are provided in Fig. S9.

For surface NO₂ predictions evaluated against EEA measurements, DACNO₂ relies primarily on emission proxies, geographic features, and multi-level meteorological variables, while temporal indicators and single-level meteorological features play a lesser role. The addition of the EEA NO₂ constraint in Phase-2 and Phase-3 increases the importance of geographic data, highlighting its value for high-resolution surface NO₂ estimation. As shown in Fig. S9, land cover emerges as the most influential single feature (36.6 %) in DACNO₂-Phase-3. Multi-level meteorological variables dominate the meteorological contribution, suggesting partial redundancy between single-level and multi-level meteorological inputs.

For NO₂ estimates by layer evaluated against CAMS, the distribution of input feature importance at lower layers (up to 1000 m) is similar to that for surface NO₂ evaluated against EEA, suggesting that DACNO₂ remains relatively stable across training phases with different constraints. Differences between the three-phase models are most apparent near the surface but gradually diminish with height. The importance of geographic features steadily decreases with height, whereas emission features reach their strongest influence at approximately 500 m before declining. Above 3000 m, both become negligible, reflecting the transition from the Planetary Boundary Layer (PBL), which is influenced by local surface features, into the free troposphere, which is dominated by broad-scale processes. In contrast, temporal indicators, single-level meteorological features, and especially multi-level meteorological features become increasingly important with height. This shift highlights the greater reliance on temporal and large-scale atmospheric information for NO₂ estimates at higher layers. Among these features, radiation flux is the most important single-level meteorological variable, and wind is the dominant variable among all meteorological features (Fig. S9). Given the consistently low overall contribution of single-level meteorological variables, future model development may consider reducing or refining the use of this feature group to streamline the input space.

Overall, the DACNO₂ model is developed by combining multi-scale inputs and multi-source constraints. The fine-scale spatial structure on the 2 km grid is primarily informed by high-resolution emission-related proxies and geographic features, whereas large-scale spatiotemporal variation and vertical structure are driven by meteorological variables and temporal indicators. Through the phased training strategy, the CAMS constraint transfers large-scale spatiotemporal variation to the DACNO₂ model, and the EEA constraint guides the model to use fine-scale static inputs to shape this variation on the 2 km grid spatially.

Figure 6 compares the average 2023 NO₂ profile estimates from DACNO₂-Phase-3 (2 km × 2 km) and CAMS (10 km × 10 km) for the Paris and Alpine regions, with results overlaid on Google Earth imagery for geographic context. In Paris, the regional average profile (Fig. 6a) indicates that DACNO₂ yields higher NO₂ concentrations up to 2000 m and steeper vertical gradients compared to CAMS. This enhancement likely results from DACNO₂'s use of high-resolution emission proxies and land cover information, allowing the model to resolve smaller and more localized emission sources (Kuik et al., 2018; Shahrokhishahraki et al., 2022). At the local scale, we take a transect over the grids of 100 km over Paris to compare the interpixel profile variability from CAMS and DACNO₂ (Fig. 6c). It is observed that DACNO₂ more clearly distinguishes spatial variability in the vertical structure, showing sharper contrasts and more pronounced local peaks than CAMS, particularly below 250 m. The regional average profiles for the Alpine area are similar between DACNO₂ and CAMS (Fig. 6b), which is due to the overall lower concentrations over this region with limited emission sources. However, local differences remain visible across a 200 km transect (Fig. 6d). DACNO₂ captures higher NO₂ concentrations around urban and small-scale hotspots, especially in valleys and canyons where pollutants tend to accumulate. Conversely, DACNO₂ provides lower NO₂ estimates in areas between the mountains with few sources. Overall, DACNO₂ provides more spatially detailed 3D NO₂ fields, revealing greater variability in the vertical profiles across different grids in this complex terrain. This refinement is important, as small point and line sources can contribute significantly NO₂ in mountainous regions (Kim et al., 2021).

To assess how vertical profiles differ between the two models across environments, we analyzed the mean DACNO₂-to-CAMS profile ratio across the entire study region in urban, suburban, rural, and uninhabited environments classified based on population density (Fig. S10) and the urbanization definition (Dijkstra et al., 2021). The results indicate that the DACNO₂-Phase-3 adjustment is not a uniform scaling of the CAMS field. Instead, near the surface, DACNO₂-Phase-3 shows higher concentrations relative to CAMS in urban regions (about 6 %) and lower concentrations in other areas (from about -20 % to -1 %). In the boundary layer (1000 m), the NO₂ concentrations are systematically higher in DACNO₂-Phase-3 compared to CAMS (from 10 % to 30 %), except in the uninhabited area (remains the same). At higher layers, DACNO₂-Phase-3 values converge to a lower ratio (about -22 % to -15 %) at 5000 m for the entire region. This behavior is also reflected in the layer-integrated column diagnostics shown in Fig. S11, which indicate near preservation of the regional column (0–5000 m), accompanied by a significant redistribution in the lower (0–1000 m) and conservative adjustment in upper (1000–5000 m) layers. Together, these results suggest that the DACNO₂-Phase-3 primarily redistributes NO₂ within the lower layers, enhancing horizontal contrast linked to human activity and emission strength, while maintaining consistently low estimates in the upper layers.

Additionally, we assessed the profile ratio across the three phases (Fig. S10). The results indicate that applying EEA constraints almost systematically increases NO₂ estimates in DACNO₂-Phase-2 relative to DACNO₂-Phase-1, likely because the CAMS data used for pretraining in Phase-1 underestimates NO₂ at EEA measurement stations. In contrast, the EEA constraint reduces NO₂ estimates in DACNO₂-Phase 3 relative to DACNO₂-Phase-2, consistent with the lower surface NO₂ levels observed in 2023 compared with the training years (2019, 2021, 2022, Fig. S7). However, the boundary-layer NO₂ estimates exhibit different trends that do not align with phase-dependent changes, which warrants further investigation.

In this work, the vertical structure of DACNO₂ is assessed through comparison with CAMS. Independent evaluation against vertically resolved observations, such as MAX-DOAS or aircraft measurements, would be the next step in future work. Such an analysis would require an hourly version of the DACNO₂ fields to provide daytime data and the application of appropriate observation operators to ensure comparability between the model and the observational data.

To assess the potential of DACNO₂ for satellite NO₂ product improvement and development, we tested its use as a source of a-priori NO₂ profiles in TROPOMI retrievals. For this, a dedicated version of the model (DACNO₂-S5P) was developed for the TROPOMI overpass time, predicting a 3 h average NO₂ (11:00–13:00 UTC) using the same three-phase strategy. The model targets, named CAMS-S5P and EEA-S5P, represent process-based and measured NO₂ data during this period.

Model evaluation (Table S1) shows DACNO₂-S5P agrees well with CAMS-S5P (RMSE = 0.98 µg m⁻³, r = 0.94, R2=0.88, bias = 0.03 µg m⁻³) on the 10 km grid. Compared to EEA-S5P measurements, DACNO₂-S5P achieves better agreement (RMSE = 5.07 µg m⁻³, r=0.77, R2=0.59, bias = 0.05 µg m⁻³) than CAMS-S5P (RMSE = 5.27 µg m⁻³, r = 0.76, R2=0.55, bias = -0.94 µg m⁻³).

We replaced the original TM5 a-priori profiles (1° × 1°, approximately 100 km × 100 km) in the TROPOMI retrievals with CAMS-S5P and DACNO₂-S5P profiles, following the approach described in Douros et al. (2023) and focusing on the troposphere. Figure 7a presents the annual mean TROPOMI NO₂ columns retrieved using these different a-priori profiles, with inter-comparisons shown in Fig. 7b and c. Both CAMS-S5P and DACNO₂-S5P profiles lead to substantial increases in the retrieved NO₂ columns, by 36.2 % and 39.8 % on average, respectively. The increase associated with CAMS-S5P is consistent with previous findings (Douros et al., 2023) and is primarily attributable to the improved spatial resolution of the a-priori profile, which better represents near-surface NO₂ enhancements and fine-scale spatial gradients, resulting in larger retrieved tropospheric columns over emission hotspots (Tack et al., 2021; Ialongo et al., 2020).

Compared to the CAMS-S5P profile, using DACNO₂-S5P as the a-priori increases retrieved NO₂ columns by about 3.0 % on average (Fig. 7b), associated with the reduced negative bias against EEA-S5P measurements reported above. This change is accompanied by a clear spatial structure in the differences, with localized increases over small-scale emission hotspots and decreases over low-emission regions. In central-western France (0–2.6° E, 45.6–46.3° N), a distinct southwest-to-northeast line of reinforced NO₂ columns appears because DACNO₂-S5P enhances the emission signals from the cities of Angoulême, Guéret, and Montluçon. Regional comparisons (Fig. 7c) show that the DACNO₂-S5P profile leads to a 1.8 % increase in the northern region and 5.9 % in Paris, with the most significant increases surrounding major emission hotspots. This is likely due to DACNO₂'s enhanced ability, based on finer resolution, to capture small-scale emission sources and resolve strong spatial gradients around NO₂ hotspots. In the Alpine region, the average increase reaches 1.7 %, ranging from -18.1 % (5th percentile) to +24.1 % (95th percentile) between the surrounding areas and the central mountains, with a similar pattern observed in the Pyrenees. The absolute difference remains small, ranging from -3.94 × 10¹⁴ molecules cm⁻² (5th percentile) to 5.39 × 10¹⁴ molecules cm⁻² (95th percentile). This large fluctuation reflects the complex NO₂ distribution in the mountainous region and benefits from high-resolution modeling, as DACNO₂ estimates can reach lower background values or enhance the hotspots signal in this region.

These results illustrate the potential of using DACNO₂ profiles to improve satellite NO₂ retrievals, particularly for evolving high-resolution instruments. However, the DACNO₂ product remains a prototype, and we outline a brief roadmap for operational deployment in the conclusion section. In addition, DACNO₂ provides a-priori NO₂ profiles up to 5000 m, while NO₂ levels at roughly 8–12 km show a slight enhancement, possibly linked to aviation and lightning (Douros et al., 2023; Kuhn et al., 2024a; Dahlmann et al., 2011; Richter, 2009). This should also be considered in future DACNO₂ development.

As noted in Sect. 2.4.1, CAMS NO₂ data for 2020 were excluded from the training set based on preliminary experiments showing that their inclusion substantially reduced model generalization at higher layers. Since 2020 was marked by the COVID-19 pandemic and large reductions in anthropogenic emissions (Levelt et al., 2022), we specifically evaluated DACNO₂'s predictive performance for this anomalous year. To this end, the DACNO₂-Phase-2 model was fine-tuned on 2020 EEA NO₂ data, following the same phased development strategy, to produce DACNO₂-Phase-3-2020.

Table 3 summarizes the 2020 evaluation results, following the format of Table 2. Both DACNO₂-Phase-2 and DACNO₂-Phase-3-2020 reproduced observed surface NO₂ concentrations well (e.g., DACNO₂-Phase-3-2020: RMSE = 5.47 µg m⁻³, r = 0.82, R2 = 0.67, bias = 0.01 µg m⁻³), with performance comparable to CAMS (RMSE = 4.99 µg m⁻³, r = 0.88, R2=0.73, bias = -2.05 µg m⁻³) but notably lower bias. Figure S5 shows the temporal trend between model estimations and EEA measurements. It is observed that DACNO₂-Phase-2 can still capture the temporal trend of NO₂ in this unknown and anomalous year, while a positive bias appears during March and May when COVID-19 control measures took place. DACNO₂-Phase-3-2020 has successfully reduced the remaining bias with the adaptive fine-tuning. This demonstrates the robustness of the DACNO₂ model and the necessity of adaptive fine-tuning to capture anomalous events. Additionally, CAMS maintains overall consistency across measurements but exhibits a pronounced negative bias, primarily in urban areas.

Agreement between DACNO₂ and CAMS remains strong at low altitudes (e.g., surface: RMSE = 2.09 µg m⁻³, r = 0.94, R2= 0.86, bias = 0.48 µg m⁻³), but declines rapidly above 1000 m, where R2 values approach zero or become negative, indicating a failure to reproduce high altitude CAMS NO₂ distributions for 2020. Comparison of CAMS NO₂ vertical distributions from 2019 to 2023 (Fig. S2) shows generally consistent annual patterns, except for 2020, which is characterized by anomalously high values above 1000 m. This anomaly is also noted in the CAMS 2020 annual evaluation report (Meleux et al., 2023), which attributes it to some sub-models producing unexpectedly high NO₂ in the upper layers, resulting in inflated tropospheric column estimates. The underlying causes remain unresolved and require further investigation. These findings highlight the importance of data screening, such as checking distributions and identifying outliers, before model training. Including biased or anomalous target data can introduce noise, increase the risk of overfitting, and reduce generalization performance.