26th Australian Institute of Physics Congress and the Australian and New Zealand Conference on Optics and Photonics

Geometry-Driven Performance Optimization in Multimode-Fiber Reservoir Computing

Not scheduled

20m

Belinda Hutchinson Building (The University of Sydney )

Contributed Oral ANZOS | Photonics and Optics (ANZCOP)

Description

Rapid growth of AI leads to an increasing demand for computational systems capable of processing large datasets. As AI workloads become large, conventional computing architectures face challenges with energy consumption and scalability. Photonic reservoir computing offers an attractive solution for certain machine learning tasks by exploiting the intrinsic dynamics of physical systems. While recent studies have demonstrated the suitability of multimode optical fibers for data processing, the influence of fiber geometry on computational performance remains unknown.
Here, we present an investigation of performance of multimode fiber-based photonic reservoir computing utilising graded-index and step-index fibers. Numerical simulations using (3+1)D beam propagation were performed for two benchmark tasks: A spiral dataset with moderate noise was employed to study the role of nonlinearity on classification performance while a subset of COVID-19 Radiography Database served as a binary image-classification benchmark.
The results reveal that graded-index fibers reach optimal performance at lower input powers compared to step-index fibers, indicating stronger nonlinear interactions. Classification accuracies >99% were obtained for spiral benchmark task showing the effectiveness of nonlinear multimode fiber dynamics. From singular value decomposition analysis of the generated feature spaces, the effective rank and spectral entropy were calculated for the COVID-19 dataset to quantify the dimensionality and information distribution of the reservoir outputs. Under similar conditions, step-index fibers generate feature spaces with substantially larger effective dimensionality and higher entropy due to stronger modal mixing. The increase in effective rank is correlated with improved classification performance, suggesting that enhanced feature-space expansion increases computation performance.
These results demonstrate that fibre geometry is an important design parameter for photonic reservoir computing as it fundamentally affects the balance between nonlinear feature generation and class separability, leading to different optimal operating regimes in graded-index and step-index multimode fibres. This understanding opens new opportunities in the development of efficient photonic machine-learning platforms.