The natural breakdown of rocks, a process called weathering, plays a key role in regulating Earth's climate by removing carbon dioxide from the atmosphere. However, measuring how much rock is weathering across a vast area like the United States is a major challenge. We studied water chemistry from rivers nationwide to trace the dissolved minerals back to their source rocks. Our analysis revealed that geological maps alone are not reliable for this purpose. For example, we often found high levels of weathering from carbonate rocks in river basins where maps showed mostly other rock types. To predict weathering across the entire country, we used a machine learning computer model. We discovered the model's predictions were significantly more accurate when it was trained using data only from natural, undisturbed environments, filtering out the 'noise' from human activity. Using this improved method, we created a new, data-driven map of rock weathering for the continental U.S. This work provides a powerful new framework for understanding a critical component of the global carbon cycle, which can help improve climate models.