Continuous vision transformer-based neural operator for scalable beam propagation modeling in optical fibers

Accurate beam propagation modeling is essential for the design and optimization of optical fiber systems. However, conventional numerical methods are often computationally expensive, while existing neural operators struggle to capture long-range evolution dynamics and preserve spatiotemporal causality. To address these limitations, we propose a scalable beam propagation neural operator (BPNO) that combines a Deep Operator Network (DeepONet) and a continuous Vision Transformer (CViT). The DeepONet learns short-range beam evolution, whereas the CViT autoregressively predicts long-range dynamics from the generated field sequences. By incorporating a three-dimensional spatiotemporal embedding mechanism, the CViT directly embeds spatiotemporal causality into its network architecture. Trained on a comprehensive dataset covering diverse fiber geometries and input conditions, the proposed BPNO achieves high prediction accuracy across multiple beam propagation scenarios. The results further demonstrate the superior capability of the CViT in modeling long-range propagation dynamics. In addition, the impacts of network hyperparameters and training data volume on model performance are systematically investigated.