Machine-learning enabled characterization of individual ring resonators in integrated photonic lattices

TL;DR

The paper tackles the challenge of extracting per-ring parameters in large coupled-ring photonic lattices from global spectral measurements. It introduces supervised learning frameworks that map spectral power distributions to onsite losses and resonance shifts, trained on data from a programmable photonic chip and validated by reconstructing experimental spectra from theory with a 1D-CNN. The two-network workflow—a regression model for $\delta_n$ and $\omega_n$ and an inverse CNN-based spectral regressor—demonstrates high fidelity parameter inference across multiple configurations, enabling scalable, non-invasive calibration and control of complex photonic circuits. This data-driven approach provides a direct link between measured spectra and the effective Hamiltonian, with potential extensions to additional couplings, larger arrays, and closed-loop optimization for automated photonic systems.

Abstract

Accurately determining the underlying physical parameters of individual elements in integrated photonics is increasingly difficult as device architectures become more complex. Inferring these parameters directly from spectral measurements of the system as a whole provides a practical alternative to traditional calibration, allowing characterization of photonic systems without relying on detailed device-specific models. Here, we introduce a supervised machine-learning strategy to learn the onsite losses and resonant frequency shifts of each individual ring in an array of coupled ring resonators from measured spectral power distributions of the whole array. The neural network infers these parameters with high accuracy across multiple experimental configurations. Our methodology provides a scalable and non-invasive method for extracting intrinsic parameters in coupled photonic platforms, paving the way for future development of automated calibration and control methods.

Figures (3)

• Figure 1: (a) Schematic of a system of $8$ coupled ring resonators with onsite losses $\delta_n$ and resonant frequencies $\omega_n$ for each $n$-th ring. (b) Schematic of the machine-learning approach. Experimental spectra from the programmable photonic platform amin2024 are used to train and test the model, the network learns to infer the intrinsic parameters like losses and resonant frequencies from measured data of an array of $8$ optical components. (c) Schematic of the inverse problem where we map simulated theory data to the expected experimental data, and compare it with the true experimental data.
• Figure 2: (a) Inferred onsite losses $\delta_n$ (red dots) versus true values (blue line) for case A, where all rings share the same resonant frequency $\omega_n$. (b) Inferred resonant frequencies versus true values for case B, where all rings share the same loss. (c) and (d) Inferred losses and resonant frequencies, respectively, versus true values, for case C, where both loss and resonant frequencies vary across rings.
• Figure 3: (a-f) show normalized spectral power $P(\Omega,n)$ as color. Panels (a–b) show spectra sorted by a function of the onsite losses $f(\delta_n)$, (c–d) by the resonant frequencies $f(\omega_n)$, and (e–f) by the joint metric $f(\delta_n,\omega_n)$. In all cases, the ordering ensures that the dominant spectral peak shifts monotonically to higher frequencies, enabling a direct visual comparison between true and learned spectra. The ML-inferred spectra reproduce the correct spectral shape, with most discrepancies limited to the absolute intensity scale.