<p>Optimization of land ecosystem carbon sink inversions based on data assimilation requires the use of an atmospheric CO₂ transport model as the observation operator to assimilate observed atmospheric CO₂ concentrations, thereby achieving posterior optimization of prior land ecosystem carbon fluxes. However, the high computational cost of atmospheric CO₂ transport models and the complexity of constructing adjoint algorithms have limited the efficiency of land carbon sink inversion, often requiring high-performance computing clusters and resulting in low assimilation efficiency. In this study, an efficient atmospheric CO₂ intelligent transport model was developed using the GraphCast deep learning network. A neural network automatic differentiation algorithm was adopted to solve the adjoint of the observation operator, and a deep learning optimization algorithm was used to minimize the four-dimensional variational assimilation (4DVAR) cost function. This approach enabled the construction of a 4DVAR algorithm within a deep learning framework for land ecosystem carbon sink estimation. Experimental results demonstrate that the proposed machine learning assimilation algorithm can achieve global carbon flux inversion at a 1° × 1°/3-h resolution for the next seven days in approximately four minutes, while maintaining reliable accuracy.</p>