<p>High-performance computing systems are increasingly driving collaborative scientific discoveries. However, in such collaborations, data is often stored in a distributed manner, is sensitive, and is strictly restricted by the policies of various institutions. Federated learning, as a paradigm for multi-party collaborative AI model training, can effectively utilize these distributed data resources without sharing the original data. However, differences in device users’ preferences and behavioral habits lead to highly heterogeneous local data, which severely limits the generalization ability and practical effectiveness of the global model. To address this challenge, we propose a personalized federated learning scheme, FedCoSim, based on cosine similarity. This method calculates the cosine similarity between the global model and each user device model in the parameter space layer by layer, measuring the directional consistency at the model level, effectively narrowing the deviation between the global and local models in the parameter space, thereby enhancing the global model’s generalization ability in heterogeneous data scenarios. Specifically, FedCoSim first evaluates the structural similarity between the global and local models layer by layer using cosine similarity and constructs a maximum similarity model based on the similarity coefficient and regularization value; then embeds this model into the local training process, reinforcing the aggregation of common knowledge while preserving the local model’s adaptability to data specificity. In addition, we provide a rigorous convergence analysis of the proposed method and conduct systematic experimental comparisons with five federated learning methods under various heterogeneous data distributions. The results show that FedCoSim significantly speeds up the convergence of the global model while maintaining model accuracy and stability, and effectively enhances its generalization ability in highly heterogeneous environments.</p>