Training datasets construction for deep learning based on core-offset model in few-mode fibers

Deep learning has become a valuable tool for analyzing and compensating phenomena in optical communication systems. However, to achieve zero-shot with no experimental samples training, it is essential to generate simulation datasets that accurately reflect the behavior of real systems under various conditions. Therefore, in this paper, we propose a novel approach for constructing simulation training datasets based on a core-offset model based on vector modes in few-mode transmission systems. This model accurately characterizes the coupling between vector modes under core offset conditions, generating optical spot distributions that closely resemble those observed in experimental environments. Through the simulation of optical spots and comparison with experimental data, we achieve an average correlation coefficient exceeding 0.80. Our research demonstrates that the core-offset model servers as a reliable tool for simulating complex optical phenomena, laying the foundation for achieving zero-shot with no experimental samples training based on deep learning models in few-mode transmission systems.