High-resolution 3D fields of nitrogen dioxide (NO2) are critical for air quality management and satellite retrievals, yet traditional chemistry-transport models (CTMs) face challenges in fine-scale modeling. Machine learning (ML) alternatives often struggle with generalization and transferability, inheriting biases from CTMs or being limited by sparse surface measurements. We present the Deep Atmospheric Chemistry NO2 model (DACNO2), a deep learning model that generates daily 2 km × 2 km 3D NO2 fields over Western Europe. The model's three-phase multi-constraint training strategy begins by pre-training on European Copernicus Atmosphere Monitoring Service (CAMS) reanalysis data to learn large-scale atmospheric patterns, then fine-tunes with CAMS and in-situ European Environmental Agency (EEA) surface data to correct biases and refine local detail, and completes with an adaptive fine-tuning to capture evolving trends. An evaluation for 2023 shows that DACNO2 reproduces broad-scale 3D CAMS fields (R2=0.90) and improves agreement with independent EEA stations over the CAMS reanalysis (R2 enhanced from 0.61 to 0.66; bias reduced from −1.15 to −0.38 µg m−3). The model resolves spatial details and exhibits physically plausible behavior. This hybrid training approach fuses the physical consistency of a process-based model with the real-world surface measurements, overcoming the limitations of using either constraint alone. Applying DACNO2 a-priori profiles to TROPOMI retrievals increases tropospheric NO2 columns by 3 % on average over those using European CAMS profiles, with enhanced contrast between low- and high-NO2 regions, primarily attributable to improved resolution. These results demonstrate the framework's potential to advance air quality monitoring and satellite remote sensing.