Measuring the Coupling Phase in Strongly Coupled Ring Resonators

A ring resonator is a closed optical loop that traps light at wavelengths where one trip around the loop adds a whole number of wavelengths of phase. Textbook design treats the point where light leaks between the ring and its feeding waveguide as a single point that splits power and does nothing else. Two theoretical papers argue that this picture fails for very thin waveguides placed very close together. In that regime the coupling region has real length, and light picks up extra phase while crossing it. The predicted consequence is a shift of the resonance by several to tens of nanometers. The same papers predict a second effect: when the surrounding liquid changes, part of the sensor's response comes from the coupling region itself and not from the loop. If this is right, refractive-index sensors built from such rings are being calibrated with the wrong physics.

Neither paper reports a single measurement. This note proposes the first experiment designed to tell the corrected resonance equation apart from the usual practice of fitting a correction factor. It rests on two lithographic controls. The first is a chip of silicon rings whose loops are identical and whose coupling regions differ only in gap and length. The second is a set of cladding windows that expose either only the coupling region or only the loop to a test liquid. Together they let us measure the coupling-induced phase, and the coupling region's contribution to sensitivity, directly instead of inferring them.

Core Claim

The modified resonance equation makes predictions that an empirical correction cannot. For a fixed gap, the extra round-trip phase should grow linearly with the length of the coupling region, with a slope set by a coupling parameter that depends on the gap. A full-vector mode solver can compute that slope from the waveguide cross-section alone. The share of sensitivity that comes from the coupling region should appear only when the liquid touches the coupling gap.

Point-coupling with a fitted per-device length offset can reproduce any single resonance position. It cannot predict how the offset changes with coupling length, gap, or the surrounding index on devices it was not fitted to. Our claim is therefore testable. With at most two or three fabrication-bias parameters calibrated on a subset of devices, the modified equation should predict held-out devices and window-isolated sensitivities better than point-coupling corrected empirically. If it does not, the theory's practical relevance is falsified at the scale it was proposed for.

What Exists and What Is Missing

The method this proposal tests comes from two papers. Ye Chen et al. derive a coupled-wave resonance condition for microfiber loop and coil resonators, written $\beta L + cMT = 2m\pi$. Here $\beta$ is the propagation constant, $L$ the loop length, $c$ a butt-coupling coefficient, $M$ a coupling parameter, and $T$ the interaction length. Using full-vector finite-element simulation, they predict $M$ as large as $-13{,}844\ \text{m}^{-1}$, resonance shifts up to $3.89$ nm, and a coupling contribution to sensitivity of up to $83$ nm/RIU, about $10\%$ of the total (RIU is a refractive index unit).

Wei Guo, Fei Xu and Yanqing Lu extend the analysis to subwavelength silicon-on-insulator photonic wire rings. They predict shifts reaching $-20.6$ nm as the wire shrinks. For a $353$ nm wire at a pitch of $1.2$ times the width, they predict a coupling sensitivity of $20.8$ nm/RIU, or $104$ nm/RIU when $\alpha=1$. They also conjecture, without derivation, that strong coupling does not degrade the quality factor (Q, the ratio of resonance wavelength to linewidth).

Both papers motivate this work and frame its hypotheses. Neither fabricates anything. Chen et al. explicitly assume that diameter and spacing fluctuations average out over long coupling lengths.

Experimental ring work in the file uses the uncorrected model. Dutta et al. build a silicon nitride ring tuned by liquid crystal and treat all tuning as phase accumulated in the cavity. Their measured loss of about $0.34$ dB/µm exceeds the simulated value because of sidewall roughness. That paper constrains our design: nominal-geometry predictions will be biased, so any credible validation has to be differential and has to absorb fabrication bias explicitly. The split-ring microwave sensor study supplies a methodological template. It calibrates resonance shift and Q change against reference liquids. Like all refractometric validations in the file, it reports sensitivity only as a single total, and that is exactly what we need to break apart.

From the broader validation literature we borrow practice, not physics. Xiao et al. report percentage discrepancies between measured and simulated bridge modes ($2.40\%$ and $3.63\%$), and they show that a sensor at a nodal point can miss a mode entirely. Both points shape our placement of replicas and our choice of error metrics.

The heavy calibration machinery in the file does not fit this problem. That includes 41-parameter Bayesian optimization in the Pebax coarse-graining study and diffusion-based multi-fidelity calibration in Diffusion-Based Surrogate Modeling. Our physical model has only a handful of unknowns. We deliberately use small least-squares fits so that the comparison stays about physics and not about calibration capacity.

Research Program

The test platform is silicon-on-insulator, chosen over microfibers because lithography controls the gap and coupling length to a few nanometers. Microfiber gaps set by hand would add uncertainty larger than the $3.89$ nm effect.

Each design block holds rings with identical loop circumference and wire width. Widths span roughly $350$–$450$ nm, bracketing Guo et al.'s $353$ nm case. Within a block, only the coupling region varies. The edge-to-edge gap runs from about $0.2$ to $1$ times the width, and the straight coupling length $T$ takes several values from point-like up to a few tens of microns. Large-gap rings in each block serve as references where the coupling phase should be negligible. Every design is repeated several times across the die so that within-chip resonance scatter becomes a measured noise floor and not a hidden confounder.

The measured quantity is a coupling phase, not an absolute wavelength. For each device we read the free spectral range (the wavelength spacing between adjacent resonances) and from it the loop group index $n_g$. We then match resonance order $m$ to the neighbouring reference ring and convert the shift into an extra round-trip phase:

$$\phi_c = \frac{2\pi n_g L\,(\lambda_{\text{ref}} - \lambda_R)}{\lambda^2},$$

where $\lambda_{\text{ref}}$ is the reference ring's resonance of the same order, $\lambda_R$ the device's resonance, and $\lambda$ the operating wavelength. The modified equation predicts $\phi_c \approx cM(g)\,T$: linear in $T$ at fixed gap $g$, with intercept zero for vanishing $T$ apart from bend contributions, and with $M(g)$ computed by the mode solver from the measured cross-section. Point-coupling predicts $\phi_c = 0$.

The second arm isolates the coupling region's sensitivity. On a second chip with identical device blocks, the oxide cladding is opened in three ways: only over the coupling region, only over the loop away from the coupler, or over everything. Each device is flushed with a series of reference liquids of known index. The coupling-only window directly measures the coupling sensitivity $S_k$. The loop-only window measures the loop sensitivity $S_n$, corrected for the fraction of the loop that is exposed. The full window tests whether $S = S_n + S_k$ holds additively.

A third arm addresses the dispute over Q. For each gap and length, we fit the all-pass transmission to separate intrinsic loss from coupling loss. The fit has an ambiguity between under- and over-coupling, which the monotonic gap series resolves. This tests Guo et al.'s conjecture that strong coupling adds no intrinsic loss, against the alternative that roughness and mode mismatch in a narrow gap add scattering of the kind Dutta et al. observed.

Fabrication bias enters the model as three parameters, each with a physical reading: a global width offset, a sidewall angle, and a gap offset. Scanning electron microscopy of sacrificial cross-sections bounds them. They are fitted on a calibration subset only.

Baselines

The do-nothing reference is the point-coupling equation with nominal effective index and loop length. It predicts zero coupling phase and attributes all sensitivity to the loop. Any measurable, gap-dependent $\phi_c$ beyond the replica noise floor rejects it. Its role is to show that the effect exists, not to compete on prediction.

The strongest simple competitor is point-coupling with empirical correction: a fitted effective length per device, or a smooth regression of the offset on gap and coupling length, using the same calibration subset. Both it and the modified model get the same three fabrication-bias parameters and the same tuning budget. This is the method practitioners actually use, and it will fit the calibration devices at least as well as the physics model.

Beating it means two things. First, lower prediction error on $\phi_c$ for held-out combinations of gap and length, especially for extrapolation to lengths outside the calibration range. Second, correct prediction of the window-isolated $S_k$, which the empirical model can only get from additional liquid-calibration data the physics model does not see. We will report error in nanometers of resonance position and as a percentage of the measured shift, in the style of Xiao et al. No reported experimental result was found on Lacuna for coupling-induced resonance shifts or isolated coupling sensitivity. The only numbers to beat are the model predictions themselves: up to $-20.6$ nm shift and $20.8$ nm/RIU for $S_k$ from Guo et al..

The strongest physics competitor is full three-dimensional electromagnetic simulation of each fabricated ring, using the same fitted bias parameters and run on a representative subset. It makes no lumped assumptions, so it should be at least as accurate.

The ablation that isolates the key ingredient is the modified equation with the coupling-length term removed. This keeps a gap-dependent but length-independent phase. If it performs as well as the full model, the distributed-coupling mechanism is not what the data support.

Why not just…?

Why not just fit an effective cavity length and move on? For one device and one liquid, that works, and we expect it to fit the calibration set well. It fails in two places. It cannot predict how the offset scales with coupling length on unseen devices without a separate fit per design. And a length correction scales with the loop's response, so it implies no sensitivity in a coupling-only window, whereas the theory predicts tens of nm/RIU there. The held-out and window experiments are built to expose exactly these failures. If the empirical correction predicts held-out devices equally well and the coupling-only window shows no response above noise, we will report that the modified equation adds nothing practical at these dimensions.

Why not just simulate the whole ring in 3D and skip the analytic equation? This is a real threat to the modelling contribution and none to the experimental one. A 3D simulation includes the coupling phase automatically. The file's argument against it is cost across large design sweeps, raised in the diffusion surrogate and Bayesian optimization papers, and that argument is about convenience, not correctness. We will therefore run 3D simulation on a subset and report its agreement with the cross-section-based equation. If the two agree within the replica noise floor, the equation earns its place as a cheap, interpretable design tool, and the separation into $S_n$ and $S_k$ becomes something a designer can engineer. If they disagree, the equation's single-mode, uniform-section assumptions are what failed, and we will say so.

Why not just update the effective index until the uncoupled equation matches, as structural engineers tune a modulus? Tuning the index shifts every resonance and every sensitivity together, so it cannot produce a response confined to the coupling window. Kim and Lee found that homogenized concrete models missed the ringing tail caused by the rebar until the bar was modelled explicitly. The coupling region is our rebar. Our gap-and-length series tests this directly: a single updated index cannot reproduce a phase that grows with $T$ at constant loop length.

Decisive Checks

The first check is whether $\phi_c$ grows linearly with coupling length at fixed gap. The slope must be resolvable above the replica scatter, and its gap dependence must fall within the uncertainty band produced by the fitted bias parameters around the solver-computed $M(g)$. A flat or noise-dominated $\phi_c$ falsifies the claim of nanometer-scale shifts.

The second check is the coupling-only window. A measured $S_k$ of the predicted order would be the first experimental isolation of the coupling region's sensitivity. A null result would bound it.

The third check is whether intrinsic loss rises as the gap narrows, which settles the Q conjecture in one direction or the other.

As a conditional extension, if $S_k$ and $S_n$ are measured with opposite signs anywhere in the explored range, we would fabricate the geometry that balances them. That would test the cancellation regime Guo et al. proposed but never built.