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

Improving inverse-designed transverse mode (de)multiplexers using machine learning interpretability

Not scheduled

20m

Belinda Hutchinson Building (The University of Sydney )

Contributed Oral ANZOS | Photonics and Optics (ANZCOP)

Description

Inverse design methodologies have shown great promise in improving on-chip nanophotonic components. Unlike conventional design techniques, which tune only a handful of basic parameters such as waveguide widths or gaps, inverse design algorithms like SPINS (Stanford Photonic INverse design Software) [1] optimise over a much larger parameter space to generate compact, unintuitive, and often better-performing designs. However, the unintuitive nature of these inverse-designed topologies makes them difficult to understand and improve. A similar lack of transparency is prevalent in many machine learning-based models, and interpretability techniques are being increasingly used to gain insights into these models. In this work, we use the machine learning-based interpretability tool LIME (Local Interpretable Model-agnostic Explanations) [2] to better understand and improve the bandwidth characteristics of inverse-designed transverse mode (de)multiplexers. A dataset of 329 SPINS-generated (de)multiplexer designs is used to obtain a LIME heatmap that highlights topological features distinct to high-bandwidth devices. Insights from LIME help us choose more effective initial conditions for SPINS optimisation leading to more broadband (de)multiplexers. All of our LIME-inspired devices showed 0.5-dB bandwidths exceeding 200 nm for both TE0 and TE1 modes compared to only 2% of devices in the original dataset. Our results [3] show that interpretability techniques can reveal underlying patterns in inverse-designed topologies as well as help improve device performance. Additionally, we extend the LIME-inspired initial conditions to more devices such as four-mode (de)multiplexers and beamsplitters, improving bandwidths in all cases while also reducing insertion loss and increasing fabrication tolerance for certain devices.

References:

1. L. Su et. al., Applied Physics Reviews 2020, 7, 1.
2. M. T. Ribeiro et. al., In Proceedings of the 22nd ACM
   SIGKDD International Conference on Knowledge Discovery and Data
   Mining, KDD ’16, 2016 1135–1144.
3. L. Pira et. al., Advanced Photonics Research 2026, 7, 5 e202500284,
   URL: https://doi.org/10.1002/adpr.202500284.