As the CO₂ observation network evolves toward higher resolution and denser coverage, monitoring anthropogenic emissions from facilities and point sources increasingly relies on physical transport models operating at fine scales. These models establish a relationship between upwind CO₂ fluxes and downwind concentration enhancements, known as CO₂ footprints. However, this conventional approach is typically computationally and storage intensive at super resolution. Advances in machine learning offer a promising solution to this challenge. We developed a machine learning model, FootNet-PS, which combines a U-Net architecture with multiple PixelShuffle layers. The model efficiently predicts the influence of upwind fluxes on downwind regions at super resolution, using only low-resolution meteorological fields, thereby significantly reducing computational and storage demands. Training data were generated using the Large Eddy Simulation application of the Weather Research and Forecasting model (WRF-LES). Evaluations show that FootNet-PS predictions align well with WRF-LES results under most conditions and the correlation coefficients are as high as 0.96. In very complex and rare meteorological scenarios, the correlation coefficients still reach 0.85. Importance validation tests using a permute-and-predict (PaP) method revealed that the 2D Gaussian plume provides the most critical information to the model, while nearby horizontal winds help capture finer structural details of the footprints. This approach holds considerable potential for rapidly emulating super-resolution footprints and mitigating computational constraints in future applications, making it possible to report near-real-time point source emissions.