Waveguide Propagation Loss by the Cutback Method

Scope

This article describes the cutback method for measuring optical propagation loss in integrated photonic waveguides. Coverage includes the destructive single-device cutback variant (traditional), the paired-device variant used in foundry process control, fit methodology, and the dominant sources of measurement uncertainty. The article applies to silicon-on-insulator (SOI), silicon nitride (SiN), and InP-platform waveguides. Loss measurement methods based on resonator linewidth (ring resonators, Fabry–Pérot fringes) are covered in a separate article.

Definition

Optical power transmitted through a length $L$ of waveguide decays as

$$P(L)=P_0\cdot {\eta }_{\text{c}}^2\cdot {10}^{-\alpha L\mathrm{/}10},$$

where $P_0$ is the input optical power, ${\eta }_{\text{c}}$ is the coupling efficiency at each end (assumed equal), and $\alpha$ is the propagation loss in dB per unit length. The factor of ${\eta }_{\text{c}}^2$ accounts for coupling loss at both the input and output facets.

Taking the logarithm:

$$10{\log }_{10}P(L)=10{\log }_{10}{P}_0+20{\log }_{10}{\eta }_{\text{c}}-\alpha L\mathrm{.}$$

Measured transmitted power in dBm versus waveguide length $L$ is linear in $L$, with slope $-\alpha$ and intercept determined by the source power and coupling efficiency.

The cutback method exploits this linearity: by measuring the transmitted power for waveguides of multiple lengths, the slope of the resulting plot yields $\alpha$ directly, without requiring an absolute measurement of either $P_0$ or ${\eta }_{\text{c}}$.

Typical propagation loss values for common integrated platforms:

Method variants

Destructive single-device cutback

In the original cutback method, transmitted power is measured for a single waveguide, the waveguide is then cleaved or polished to a shorter length, and the measurement is repeated. The process is iterated for typically four to six lengths. Slope of $P(L)$ in dB versus $L$ gives $\alpha$.

The destructive cutback is the most accurate variant in principle because the input coupling geometry remains nearly identical across measurements (the same fiber, same alignment, same launch conditions). In practice, several issues limit accuracy:

• Each cleave or polish step introduces facet quality variation
• Fiber realignment after each cut introduces coupling efficiency variation
• The measurement is destructive and cannot be repeated on the same chip
• Most chips do not have sufficient length for multiple cuts

Use of the destructive variant is now largely restricted to optical fiber loss measurement (where meters of length are available) and to specialty research devices.

Paired-device (multi-length) cutback

In the paired-device variant, a set of nominally identical waveguides of different lengths $L_1,L_2,L_3,\dots$ is fabricated on the same chip. Transmitted power is measured for each waveguide using identical coupling geometry. Slope of $P(L)$ in dB versus $L$ gives $\alpha$.

This is the dominant variant in modern integrated photonics. Foundry process control modules (PCMs) and standard cell libraries (such as those provided by AIM Photonics, IMEC, and GlobalFoundries) include cutback waveguide sets specifically for this purpose. Typical set: 4–8 waveguides spanning $L_{\text{min}}=0.1$ cm to $L_{\text{max}}=10$ cm in a spiral or boustrophedon layout.

The paired-device variant trades the perfect-coupling assumption of destructive cutback (which is itself imperfect in practice) for a multiple-device assumption: that fabrication variations across the multi-waveguide set are negligible compared to the propagation loss signal. For modern foundries with well-controlled processes, this trade is favorable.

Spiral waveguide variants

For very low-loss platforms (silicon nitride at $<0.1$ dB/cm), accumulated loss over centimeter-scale waveguides is often below the measurement noise floor. To increase sensitivity, individual cutback waveguides are routed as long spirals, with the longest path being meters rather than centimeters. The shortest waveguide in the set must use the same bend geometry as the longest to ensure bend-loss contributions cancel.

The Archimedean spiral and double-spiral (no internal bend cross-overs) are common layouts.

Equipment

For SOI and silicon nitride platforms, grating coupler arrays at standard pitch (typically 127 μm or 250 μm) enable measurement of all waveguides in the set using a single fiber array, eliminating the need for re-alignment between waveguides.

Procedure

1. Layout verification

Confirm the waveguide lengths in the test set against the layout file or PDK documentation. Reported lengths must include all bends, transitions, and any additional structures in the optical path. Inconsistent length reporting (e.g., reporting straight-section length but excluding bend length) is a common source of large systematic error.

2. Establish reference alignment

Align fibers to the input and output couplers of one waveguide in the set, typically the shortest. Optimize coupling to peak transmission. Record the alignment offsets.

3. Measure transmission for each waveguide

For each waveguide $i$ in the set, translate to the corresponding coupler pair without re-running first-light search. For a properly designed PCM with standard-pitch couplers, lateral translation alone is sufficient. Record the transmitted power $P_i$ in dBm at the design wavelength.

If wavelength-resolved measurements are required (for spectral characterization of loss), record the full spectrum at each waveguide using a swept tunable laser. The propagation loss is then extracted independently at each wavelength.

4. Verify coupling-loss consistency

The coupling loss at each end of each waveguide must be identical within tolerance. Two checks identify violations:

• Visual inspection: confirm all grating couplers in the set are nominally identical (same design, same orientation, same nominal coupling angle).
• Repeat measurement on the shortest waveguide after measuring the longest; the change in transmitted power must be $<0.1$ dB.

Drift larger than 0.1 dB during the measurement set indicates either fiber position drift, source power drift, or temperature drift. Stabilize and repeat.

5. Fit

Plot transmitted power in dB versus waveguide length $L$. Perform a least-squares linear fit:

$$P_i\text{ [dB]}=-\alpha \cdot L_i+b,$$

where the slope $-\alpha$ is the propagation loss and the intercept $b$ contains the source power and the coupling term $20{\log }_{10}{\eta }_{\text{c}}$ for both couplers.

Report:

• $\alpha$ in dB/cm
• Fit uncertainty ${\sigma }_{\alpha }$ from the linear regression
• Number of waveguides in the set $N$
• Length range $[L_{\text{min}},L_{\text{max}}]$
• $R^2$ of the fit

6. Extract coupling loss (optional)

The fit intercept $b$ gives the source power minus the total coupling loss. Subtracting the source power (separately measured by direct connection of source fiber to detector) gives the total coupling loss of both couplers, $-20{\log }_{10}{\eta }_{\text{c}}$ in dB. Dividing by two yields the per-coupler coupling loss.

This per-coupler coupling loss extraction is a side-benefit of the cutback measurement and is the standard method for characterizing grating coupler efficiency in fabricated devices.

Worked example

A silicon nitride PCM contains four spiral waveguides on a single chip, with lengths $L=1,5,25,100$ cm. Transmission at 1550 nm is measured via grating couplers at standard 127 μm pitch:

Linear least-squares fit on $(L,P)$:

$$\begin{array}{c}{\displaystyle \alpha =0.069\text{ dB/cm},b=-15.3\text{ dBm},}\\ {\displaystyle R^2>\mathrm{0.999.}}\end{array}$$

The 95% confidence interval on $\alpha$ from the fit is $\pm 0.002$ dB/cm. The intercept $b=-15.3$ dBm, compared to a source power of $+0$ dBm, yields a total coupling loss of $15.3$ dB, or $\sim 7.6$ dB per coupler if the fiber path itself is lossless. SiN grating couplers, uniform or apodized, typically lose 4 to 6 dB at 1550 nm, so the excess points to misalignment, a polarization mismatch or an unoptimized grating; the slope, and so $\alpha$, does not depend on it.

The extracted $\alpha =0.07$ dB/cm is consistent with the SiN platform's specified low-loss performance.

Sources of uncertainty

Length reporting error. As above, the dominant systematic error. A 10% error in waveguide length produces a 10% error in $\alpha$. Layout-extracted lengths must be used, not nominal mask dimensions.

Coupling variation across the set. The cutback method assumes identical coupling efficiency for all waveguides. Differences in coupler fabrication, fiber position, or coupler-to-fiber angle introduce per-device variation that appears in the fit as scatter and biases the slope. For high-quality PCMs and standardized fiber arrays, this variation is $\sim 0.5$ dB; for poorly controlled measurements it can exceed 2 dB.

Source power drift during the measurement. Tunable lasers typically drift by 0.01–0.05 dB over minutes. For long measurement sets, this drift introduces apparent trends in the data. Use of a reference-tap with a second power meter that monitors source power throughout the measurement and corrects for drift is standard for high-precision measurements.

Polarization drift. For polarization-sensitive waveguides (TE-only grating couplers, polarization-maintaining waveguides), small polarization drift produces 0.1–1 dB per-measurement variation. Use of polarization-maintaining input fibers and active polarization stabilization eliminates this contribution.

Bend losses included in extracted $\alpha$. Real PCM waveguides include bends. If bend density is the same across the cutback set (proportional to length), bend loss contributions appear as a fixed contribution to $\alpha$ and the extracted value is the total loss per unit length (bend plus straight). If bend density varies across the set, the extraction is invalid. For accurate straight-waveguide $\alpha$ alone, the bend loss must be characterized separately.

Insufficient length range. Linear regression on a small length range $[L_{\text{min}},L_{\text{max}}]$ with $L_{\text{max}}-L_{\text{min}}$ comparable to a few measurement uncertainties yields high fractional uncertainty on $\alpha$. Standard guideline: $\alpha \cdot (L_{\text{max}}-L_{\text{min}})>10{\sigma }_P$, where ${\sigma }_P$ is the per-measurement power uncertainty. For low-loss platforms this drives the use of spirals and long total path lengths.

Substrate leakage. For SOI waveguides the lower oxide cladding has finite thickness (typically 2 μm), and a small fraction of the optical field can leak into the silicon substrate. The leakage is the same in every centimeter, so it adds to the fitted $\alpha$ rather than bending the line; at 2 μm of oxide it is negligible for TE and matters for TM, thinner oxides and modes near cutoff (see substrate leakage).

Validation

The fit residuals should be randomly distributed about zero. A systematic trend in residuals indicates stray light or cladding modes reaching the detector (the longest, weakest structures then read high), bends or widths that do not scale with length across the set, polarization drift, or scattering at structural defects, none of which the constant-$\alpha$ model captures.

The extracted $\alpha$ should fall within the expected range for the platform (table above). Values dramatically inconsistent with the platform suggest either a process excursion, a polarization mismatch, or an extraction error.

For paired-device cutback, repeated measurement of the shortest waveguide at the end of the run should agree with the first measurement within 0.1 dB. Disagreement indicates drift and the run should be repeated.

How cutback compares with the Fabry-Perot fringe and ring-Q methods, with worked uncertainties for each and a rule for which to use at what loss level, is in Three ways to measure waveguide loss, compared.

References

For the original cutback method as applied to optical fibers, see ITU-T G.650.1. For modern cutback procedures in integrated photonic foundry process control, see the AIM Photonics PDK documentation and the IMEC iSiPP50G manual. For an analysis of cutback method uncertainty and comparison to ring-resonator-based loss extraction, see Vermeulen et al. (2016) on add-drop ring resonator loss characterization.