Abstract
<title>Abstract</title> <p>High-resolution atmospheric simulations including the meteorological variables, trace gases, and aerosols are essential for weather, climate, and air quality assessments. However, systematic observations are often limited in developing regions, while the high computational cost of conventional chemistry-transport models restricts their frequent applications. Here, we design a framework for emulating the atmospheric variations over South Asia, computed from regional model - Weather Research and Forecasting model coupled with Chemistry (WRF-Chem). Four different DL architectures: convolutional neural network (CNN), convolutional long short-term memory network (ConvLSTM), and two U-Net configurations (3blk, 4blk) have been trained with > 1000 WRF-Chem frames for the summer conditions. The framework uses global fields of meteorology (ERA5) and chemistry (CAM-Chem) and targets to emulate WRF-Chem fields of temperature (T2m), specific humidity (Q2m), surface ozone (O3), and fine particulate matter (PM2.5), within a unified framework. Across the architectures, the emulators tend to capture the meteorological fields efficiently (SSIM and r2>0.9; NRMSE<0.1) and to an extent the O3 variability (r2 up to 0.72; NRMSE<0.2). However, the skills in reproducing PM2.5 variability were lower and spatially heterogeneous, nevertheless, comparatively better performance was noted over the Gangetic Plain (SSIM and r2 > 0.6). Among the explored configurations, the UNet-3blk with a 6-h temporal lag showed the optimal performance. The study shows the strong potential of DL emulators for performing computationally-efficient high-resolution atmospheric simulations. Extending the framework to have additional variables, and training over longer datasets from contrasting seasons are key future directions to support wider scientific analyses and policy-relevant applications.</p>