- The paper demonstrates a compact architecture for simultaneous high-precision measurement of wavelength and full Stokes polarization in optical fields.
- It employs a multimode and multicore fiber system with optical delay lines and a ResMLP neural network to decode downsampled temporal speckle data.
- Experimental results show a 0.25 pm spectral MAE and robust polarization resolution, highlighting the system’s potential for real-time fiber-optic sensing.
High-Speed Multi-Dimensional Optical Field Measurement via MMF-MCF Spatial-Temporal Mapping
Introduction
This work presents a highly compact, all-fiber architecture for real-time simultaneous measurement of wavelength and full-Stokes state of polarization (SOP) in optical fields. Leveraging a spatial-temporal mapping system comprising multimode fiber (MMF), multicore fiber (MCF), and optical delay lines, the system encodes high-dimensional optical information into minimal channels, enabling single-pixel, high-speed, and robust optical field analysis. The architecture employs a data-driven neural demodulation backend based on a residual multilayer perceptron (ResMLP) to disambiguate spectral and polarization parameters from severely downsampled temporal signatures.
System Architecture and Measurement Principle
The system exploits the intermodal interference in MMFs to generate high-dimensional speckle patterns at the output facet, which encode both input wavelength and SOP. A seven-core MCF samples the output, forming a low-dimensional representation of the speckle field. Each MCF channel passes through distinct fiber delay lines, imparting fixed temporal offsets and producing a serial 7-pulse sequence that aggregates the spatial intensity information into time. This mapping allows acquisition via a single high-speed photodetector, circumventing the speed limitations imposed by image sensors.
The ResMLP neural network is tasked with inverting the nonlinear mapping from 1×7 time-multiplexed pulses to the input optical parameters. The architecture utilizes high-dimensional projections and deep residual blocks for nonlinear regression, optimized with weighted MSE loss to preferentially minimize wavelength estimation error over SOP accuracy.
Experimental Validation
Comprehensive dataset acquisition was carried out across 40 wavelengths and up to 64 SOPs, covering step sizes from sub-degree micro-perturbations to large-scale polarization rotations. Experimental results establish the following key performance metrics:
- Spectral mean absolute error (MAE) of 0.25 pm in the finest sampling regime
- Polarization resolution of 0.2015 (mean Euclidean distance in normalized Stokes space)
- Measurement bandwidth: up to 800 pm
- Repetition rate: 100 MHz for time-multiplexed data acquisition
The system achieves these results despite utilizing only 7 spatial samples (6.3% sampling ratio of the MMF near-field), demonstrating that much of the essential information for demodulation is retained post-downsampling and can be robustly decoded with ML models.
Trade-offs: Spatial Sampling Density vs. Measurement Rate
A systematic reduction in the number of active MCF cores revealed a rapid degradation in performance below 5-6 channels. The network exhibits asymptotic improvement with increased channels, but practical design must balance between increasing spatial sampling—and thus accuracy—and preserving measurement rate, which is inversely related to the extended temporal acquisition window required for more delay lines. These findings concretely quantify the minimum spatial sampling density required for reliable, high-speed multiparameter demodulation.
Robustness and Fault Tolerance
Single-core ablation experiments demonstrate isotropic fault tolerance: removal of any individual core does not appreciably degrade spectral or SOP demodulation accuracy. This strongly supports the conclusion that the high-dimensional optical information is redundantly distributed across the MMF's speckle field. This property relaxes constraints on MMF-MCF fusion and alignment, simplifies practical packaging, and enhances the reliability of field-deployed systems.
Implications and Future Directions
From a practical perspective, this architecture marries high-speed operation, high-precision multi-parameter optical analysis, and minimal hardware complexity. The robust operation under partial hardware failure and redundancy in information encoding also make the solution attractive for real-world, maintenance-limited environments. This approach directly addresses the bottlenecks of image sensor-based field analyzers, which are hampered by data readout latencies and limited scalability.
Theoretically, the results corroborate that physically-inspired sparse sampling—combined with expressive nonlinear models—enables efficient inversion and decoupling of multiplexed physical phenomena even under strong measurement constraints. This insight is likely generalizable to other computational sensing applications where high-dimensional physical states are encoded in complex, distributed patterns.
For future extensions, the architecture could support further parameter scaling via increased MCF core counts, exploration of alternative mode-multiplexing, and advanced neural network architectures tailored for nonlinear signal demultiplexing under hardware constraints. In optical communications and sensing contexts, embedding such analyzers could enable real-time, adaptive, feedback-driven systems over significant deployment scales.
Conclusion
The MMF-MCF spatiotemporal mapping architecture constitutes a significant advancement in high-speed, multi-parameter optical field measurement. By leveraging strong physical encoding and data-driven nonlinear decoding, the system achieves picometer spectral precision and robust SOP estimation with minimal hardware, underlining new design principles for next-generation computational fiber-optic sensors and analyzers (2606.06841).