Towards scalable and accurate property prediction for photonic crystal fibers with federated learning.
- Publisher:
- Optica Publishing Group
- Publication Type:
- Journal Article
- Citation:
- Opt Express, 2025, 33, (16), pp. 33421-33434
- Issue Date:
- 2025-08-11
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
The inverse design of photonic crystal fibers (PCFs) is essential for optimizing fiber optic devices. Although neural network-based methods have shown promise in predicting optical properties from structural parameters, most existing models are tailored to single PCF geometries and lack generalization across diverse structures. To overcome this limitation, we propose a federated learning (FL) framework for predicting optical properties across multiple PCF types. Our approach aggregates data from structurally diverse PCFs and explores two settings: (1) a heterogeneous FL scenario where each client holds data from a distinct PCF geometry, and (2) a non-IID scenario where data from a unified PCF structure is randomly distributed among clients. To handle variations in input/output dimensions across structures, we design a flexible architecture featuring a shared hidden layer with personalized input/output layers, coupled with a dynamic padding strategy to enhance model adaptability. Experiments demonstrate that our method achieves high prediction accuracy (Setting 1: MSE < 0.0085; Setting 2: MSE < 0.0057), with strong convergence and robustness under non-IID conditions. This framework effectively addresses structural diversity and data heterogeneity, mitigating distribution drift and breaking the "data island" limitations of centralized approaches, thereby enabling generalized optical property prediction for a wide range of PCF designs.
Please use this identifier to cite or link to this item: