Real-time, spatially dense networks of sensors measuring CO 2 concentrations hold great promise for obtaining fine-grained attribution of the location, sectors, and trends of CO 2 emissions in cities. Without these sensors, emissions are tracked via emissions inventories, which are expensive to maintain and based on assumptions such as accurate reporting by emitters. Treating the emissions inventory as the Bayesian prior updated by sensor observations grounds the posterior emissions in real measurements. However, this method is prior-dependent, computationally expensive, and impractical where emissions inventories are unreliable or unavailable. We present a novel benchmark dataset for machine learning-based CO 2 emissions mapping from sparse ground-based sensor networks, addressing a gap in supervised learning resources for atmospheric inverse modeling. Our dataset consists of 1 km resolution CO 2 concentration simulations over CONUS using WRF-Chem paired with Vulcan v3.0 emission fields, structured as city-sized patches (48 km × 48 km) with 72-hour temporal windows and augmented with spatial masking to simulate realistic sensor coverage (5-100% pixel coverage). We evaluate baseline models including linear regression, XGBoost, and a modified U-Net architecture. The U-Net significantly outperforms tabular models and maintains the ability to identify large point sources at 50% coverage but losing fine spatial detail below 25% coverage. Validation on real sensor data from Houston and the Bay Area reveals challenges in translating from simulated to real-world conditions, with 10-20 sensors in each city providing insufficient coverage for detailed emission mapping, highlighting the need for denser networks and improved models to enable practical urban emissions monitoring from sparse observations.