Accurate prediction of reservoir parameters is the core of reservoir evaluation and development plan optimization, but traditional methods are difficult to effectively integrate multi-source heterogeneous data, depict complex spatial heterogeneity and ensure physical consistency. Therefore, this paper proposes a multimodal transformer graph neural operator physical constraint network (MT-GNO-PCN) to realize the joint high-precision prediction of reservoir parameters. Firstly, the multimodal transformer is used to integrate seismic attributes, logging curves and geological interpretation data to construct a unified semantic feature representation; Then the map neural operator is used to learn the continuous space mapping function and flexibly model the distribution law of reservoir parameters in irregular geometry and complex geological structure; Finally, the physical constraint loss term based on the relationship between Darcy's law and rock physics is introduced to enhance the physical rationality and generalization ability of the prediction results. Experiments on real and synthetic reservoir data sets show that the average mean square error of this method is about 46% lower than that of the traditional convolutional neural network and 33% lower than that of the model using only multimodal transformer in the prediction of porosity, permeability and water saturation; The average determination coefficient (R²) is 0.89, and the error increase is controlled within 165% under 15% noise interference, which is significantly better than the existing comparison model. The framework provides a new way for reservoir intelligent modeling driven by multi-source data with high accuracy, strong robustness and physical consistency.