Scientists often choose between two types of models to track water and carbon on land. Traditional physics‑based models use known natural laws and offer clear explanations, but they can’t always make use of the latest satellite and ground measurements—and different models can give quite different answers. Machine learning models learn directly from data and can spot complex patterns, yet they sometimes produce results that conflict with basic physical rules and struggle when conditions change.

Our model (H2CM) overcomes these hurdles by marrying both approaches into one system. It runs every day on a global grid and follows key water stores—like soil moisture, groundwater, and snow—and flows, such as how much water evaporates or runs off. At the same time, it tracks carbon uptake by plants and carbon release through respiration. By feeding in real observations—from satellite measurements of vegetation to on‑the‑ground carbon exchange data—H2CM stays grounded in reality while still learning from new information.

H2CM outperforms standalone physics models and pure machine learning, especially when it comes to seasonal carbon patterns in places like the Amazon and Southern Africa. It even picks up short‑lived bursts of carbon release after rain in dry regions—something traditional models often miss.