Abstract: Accurately predicting gross primary productivity (GPP) is crucial for understanding carbon cycling; however, most studies have predominantly investigated GPP using only environmental metrics, overlooking the pivotal role of functional traits as intermediaries between the environment and GPP and the predictive potential of GPP. Therefore, this study developed a three-dimensional “engine” framework to predict GPP and tested it by leveraging functional traits from 2,040 plant communities on the Tibetan Plateau, incorporating environmental factors and the length of the plant-growing season. Our results highlight that while the environment exerts a dominant direct influence on GPP dynamics, the contribution of leaf density traits to GPP prediction should not be overlooked. The proposed framework achieved a prediction accuracy close to 0.92, underscoring its feasibility in GPP prediction. However, incorporating the nitrogen-to-phosphorus ratio into the framework diminished the model's predictive accuracy. Within the stoichiometric dimension alone, the prediction accuracy significantly increased with the number of input traits, indicating a substantial potential for enhancing predictive capability. In the dimension of environmental factors, incorporating more environmental factors does not significantly enhance the model's predictive ability. Our research facilitates the dynamic, continuous, and relatively accurate monitoring of GPP, contributing to a better understanding of carbon cycle dynamics and supporting informed ecosystem planning and management.

Key words: carbon cycle, climate change, machine learning, productivity prediction, stoichiometry, Tibetan Plateau