Quantifying greenhouse gas emissions from atmospheric observations requires simulating how gases move through the atmosphere, a computationally intensive process that hinders timely emission estimates. We developed a machine learning-based emulator that replicates these simulations at high resolution but runs hundreds of times faster. This enables near-real-time estimation of emissions, combining diverse data sources such as satellite and surface measurements. Our approach improves both speed and accuracy, supporting scalable, timely monitoring of greenhouse gases critical for informing climate mitigation policies and verifying emissions reductions.