Machine learning has been widely used in ecosystem research. This study analyzed the dynamic variation characteristics of the carbon flux (NEE) data from the Larix gmelinii ecosystem, observed from January 1, 2014, to December 31, 2018, and simulated the data using various machine learning methods. The results indicated that: (1) The daily dynamics of the NEE during the growing season of the Larix gmelinii ecosystem exhibited a "U" shape, with the ecosystem acting as a carbon sink overall. The carbon sink capacity was strongest in July, with a monthly average of 67.57 g C m-2. From September to May of the following year, the ecosystem acted as a carbon source. (2) Structural equation modeling revealed that the main influencing factors of NEE in the Larix gmelinii ecosystem are latent heat flux (LE), net radiation (Rn), leaf area index (LAI), air temperature (Ta), relative humidity (RH), vapor pressure deficit (VPD), and soil water content (SWC), with latent heat flux and net radiation being the most dominant factors affecting NEE variations. (3) Four machine learning methods (RF, XGBoost, SVM and ANN) accurately simulated the NEE of the Larix gmelinii ecosystem, with XGBoost and RF providing similar results. However, XGBoost outperformed RF in terms of simulation accuracy and computational efficiency. The study provides a basis for using machine learning methods to estimate ecosystem carbon fluxes.