WorksPhysics and Astronomy
Physics-Informed Surrogate for Sb2S3 Dispersion Trimming in Thin-Film Lithium Niobate
Abstract
A film-thickness error of 8 nm, within the tolerance of commercial thin-film lithium niobate wafers, shifts the group-velocity dispersion at 1550 nm by about 28 ps km. We study numerically whether a segmented, non-volatile phase-change cladding can correct this error after fabrication. In mode solves of 3012 geometries, SbS switches by a median of 181 ps km at a 40 nm strip thickness with no material absorption under the adopted optical constants, whereas SbSe reaches a comparable contrast at a median modal loss of 5.6 dB cm. Because phase adds along the guide, four independently switched binary segments reach a two-dimensional region of the plane; eigenmode expansion supports this description at the junctions. Full-wave solves at 65 die geometries show that the segment states move with the die error, so the bit pattern is chosen from a per-die model: a spectral surrogate constrained by exact identities of the lossless Maxwell problem. Its accuracy within the training domain is similar to that of unconstrained baselines; it keeps and consistent, and its error at an unseen film thickness is 8.3 rather than 13.0 ps km. With a bare test ring and a second ring carrying a crystalline strip already on the mask, 96.8% of the dies are trimmed into ps km, scored on full-wave solutions. A thermal model gives a candidate write window for a 532 nm beam.
Introduction
Thin-film LiNbO combines low propagation loss [1] with a large second-order nonlinearity, a usable Kerr coefficient and a fast electro-optic response [2, 3, 4]. Its nonlinear devices, among them octave-spanning soliton microcombs [5, 6, 7], two-octave supercontinuum [8], mid-infrared generation [9] and broadband parametric amplification [10], are designed around the group-velocity dispersion and its slope . In a sub-micrometre ridge the waveguide and material contributions to are comparable and opposite in sign, so these terms are specified to the nanometre.
A fabricated die rarely meets that specification. In the sweep reported below, at 1550 nm changes by 3.6 ps km per nanometre of film thickness, so the nm thickness variation of a commercial wafer moves it by about 28 ps km, more than the ps km band used here as the target. Ridge width matters about ten times less per nanometre, so the dominant error is set by the wafer rather than by lithography. Fabrication-tolerant designs [11] lower the sensitivity at the cost of design freedom, but cannot correct a die once it is made.
Post-fabrication trimming has been shown by localized annealing [12, 13], femtosecond-laser writing [14] and non-volatile phase-change claddings [15, 16]. The last hold their state without power, and the low-loss chalcogenides SbS and SbSe [17, 18] already support multi-bit programmable circuits [19, 20] and foundry integration [21]. These schemes trim a scalar, a phase or an effective index at one wavelength. Trimming dispersion is harder for two reasons. The target has two components, and , which one continuous control cannot set independently. The trimming elements also sit on the same mis-sized ridge, so their own effect shifts with the error they are meant to correct. We are not aware of a report in which group-velocity dispersion is corrected after fabrication.
Here we study, by simulation, post-fabrication trimming of dispersion in an X-cut thin-film LiNbO ridge with a segmented SbS cladding (Figure 1 and Figure 2), and address both difficulties. Because phase adds along the guide, independently switched binary segments give two-dimensional control of . Mode solves at 65 die geometries show that the segment states do not shift rigidly with the die error, which limits a rule based on the nominal library to 83% of these dies even when the error is known exactly. We therefore choose the bit pattern from a per-die model, a spectral surrogate constrained by exact identities of the lossless Maxwell problem, and infer the die error from two test rings. One of them carries a crystalline strip already on the mask, which removes most of the near-degeneracy between film and width errors. Scored on full-wave solutions, this procedure brings 96.8% of the dies within ps km. We also report its sensitivity to a systematic solver offset and a thermal estimate of the laser write; numerical details are in the supplementary material.

Figure 1. Workflow of the study. (a) Full-wave data: 1081 mode solves with raw , from the space-filling (B1) and jitter (B2) sets of Sec. II-D. (b) Exact identities of Maxwell’s equations imposed on the surrogate (§ 5). (c) The physics-informed surrogate and the per-die loop: test rings, inferred die error and the choice of 1 of 16 bit patterns, with gradient design of the segment library as a side branch. (d) The eigenmode-expansion check of the junctions, the success rate scored on the B2 dies, and the thermal estimate of the laser write.

Figure 2. Waveguide and trimmer. (a) Cross-section of the X-cut LiNbO ridge with the SbS strip, to scale at , , and nm, with the swept dimensions labelled. Contours mark and of the quasi-TE mode at 1550 nm, strip amorphous (dashed) and crystalline (solid). (b) The four-segment trimmer of Sec. VI-A (, 600/40, 900/40 and 900/80 nm) in bit pattern 1010 (1: crystalline). Segment lengths are not to scale.
Device, Material Models and Numerical Methods
Waveguide and materials
The waveguide is a fully etched X-cut LiNbO ridge on silica, with a phase-change strip on the top surface and above (Figure 2). The optic axis lies in the film plane, so the dominant field component of the quasi-TE fundamental mode lies along the extraordinary axis. Four dimensions are swept: the film thickness (350 to 500 nm in 50 nm steps), the ridge width (900 to 1500 nm in 100 nm steps), and the strip width (300, 600 and 900 nm) and thickness . The strip is no wider than the ridge, so the mode is guided by the ridge and perturbed by the strip. Because the two materials differ in index contrast, is 20, 40, 80 and 120 nm for SbS and 10, 20, 40 and 60 nm for SbSe. A thicker strip gives a larger switching contrast but pulls the mode out of the LiNbO and, for SbSe, raises the loss.
SbS and SbSe switch reversibly between amorphous and crystalline phases with a large index contrast and, unlike the tellurides, without a strong absorption penalty in the telecom band [17]. The LiNbO film uses the extraordinary-index Sellmeier fit of Zelmon et al. [22], and the phase-change layers use the ellipsometry and of Delaney et al. [17] for the two end states (Fig. S1 in the supplementary material). At 1550 nm the SbS index rises from 2.72 to 3.31 on crystallisation and the SbSe index from 3.29 to 4.05. For SbS, is zero throughout the band in both states; crystalline SbSe has from at 1400 nm to at 1550 nm, the origin of the SbSe losses reported here. Measured SbS constants vary between laboratories with deposition conditions and stoichiometry [23]; another source would shift the absolute dispersion values rather than the trends.
Intermediate crystalline fractions follow the Lorentz–Lorenz effective-medium rule,
applied to the complex permittivities and of the end states at , 0.25, 0.5, 0.75 and 1. The device of Sec. IV uses only the end states; the intermediate points show how non-linear the response is between them.
Mode solver, dispersion extraction and the numerical floor
Modes were computed with the finite-difference mode solver in Tidy3D 2.11.2 [24] on a domain at 32 grid steps per wavelength, with every layer boundary on a cell edge (Fig. S2). Each geometry was solved at 11 frequencies spaced uniformly in across 1380 to 1720 nm, the 1400 to 1700 nm reporting band extended by 20 nm at each end. A polynomial in was fitted to and differentiated analytically to give and , with the order raised only while the fit residual continued to halve.
Each dispersive medium is fitted to a pole-residue model before the solve, and a small RMS error in the index can be a large error in its curvature: in a diagnostic run with a tolerance of , the bulk LiNbO at 1550 nm came out at 373 ps km, against 99.5 ps km at and 99.2 ps km from a spline through the raw data (Table S1). Every production solve uses the tolerance.
Convergence was assessed at a geometry on a zero crossing, where only an absolute criterion is meaningful (Table S2). Shrinking the domain or lowering the resolution moves by 41.9 and 56.1 ps km, whereas enlarging the domain moves it by 0.5 ps km, so the settings used are the smallest that are converged. A re-check on the second dataset against a finer reference keeps every amorphous variation within 4 ps km and brings the crystalline 900/80 nm strip within 1 ps km once the strip is resolved by twice as many cells (Sec. S2). We take the numerical floor as the larger of ps km and 2% of , and quote no zero-dispersion wavelength to better than nm. As a property of the discretisation, this residual is expected to be largely systematic, correlated between neighbouring geometries on the same mesh rather than random from die to die; Sec. VI-F treats it as a common offset on every die.
Dataset and quality filter
A solved geometry enters the dataset only if a guided quasi-TE mode exists, if falls monotonically with wavelength across the extended band, which rejects anticrossings with a strip-guided mode, and if the fit residual is at most . The sweep produced 3012 converged solutions, 1546 for SbS and 1466 for SbSe, covering 92.0% and 87.3% of their grids (Table S3). The rejections concentrate at the thickest strip, and no reported result rests on an excluded node.
A second, full-wave dataset for the surrogate and the die-level checks
The sweep stores but not the raw , and it samples a fixed nominal ridge, so it cannot show how the segment states move when the ridge is mis-sized. A second set of solves with the same solver, materials and quality filter keeps the raw , and TE fraction at 21 frequencies across 1380 to 1720 nm (Fig. S3). Its space-filling part draws 256 geometries from a scrambled Sobol sequence [25] over to 500 nm, to 1400 nm, to 900 nm () and to 120 nm, each solved with the strip amorphous and crystalline. Its jitter part re-solves the four mask segments of Sec. VI-A in both states, and the bare ridge, at 65 die geometries: the nominal ridge and 64 Sobol points in the box , with nm and nm. Every value used in Sec. VI-C is a mode solve at the die’s own geometry. Of the 1097 solves, 1081 passed the filter: all 585 jitter solves and all but 16 of the 512 space-filling solves.
Every solve also stores, at each frequency, the confinement factors of Eq. (5) for the extraordinary and ordinary components in the LiNbO, the strip, the silica and the air. At 1550 nm, ranges from 0.45 to 0.92, from 0.06 to 0.13 and up to 0.53. The group index rebuilt from them as the energy velocity, , matches the fitted group index to a median of .
Dispersion Tuning and Material Selection
Dispersion of the as-deposited waveguide
With the strip amorphous, at 1550 nm is normal for thin films and anomalous for thick ones. Over the SbS sweep the median value falls from +419 ps km at nm to +130 at 400 nm, −63 at 450 nm and −202 ps km at 500 nm. The zero crossing lies between 400 and 450 nm for most widths, and its locus runs diagonally across the (, ) plane (Fig. S4): a thicker film can be brought back to zero dispersion by a narrower ridge.
Between and 450 nm the median slope is −3.6 ps km per nanometre (interquartile range −4.0 to −3.1), and the median slope in ridge width is 0.56 ps km per nanometre. An 8 nm thickness error therefore moves by about 28 ps km and a 15 nm width error by about 8 ps km.
Widening the strip from 300 to 900 nm moves the zero locus toward thinner phase-change layers (Fig. S5). Counting every crystalline fraction, a zero-dispersion wavelength inside 1400 to 1700 nm occurs for 16.8% of SbS solutions and 18.9% of SbSe solutions (21.9% and 28.2% as deposited).
| SbS | SbSe | |
| Index at 1550 nm, amorphous to crystalline | 2.72 to 3.31 | 3.29 to 4.05 |
| Strip thicknesses swept (nm) | 20, 40, 80, 120 | 10, 20, 40, 60 |
| Converged geometries (grid coverage) | 1546 (92.0%) | 1466 (87.3%) |
| Solutions with an in-band zero crossing, all (as deposited) | 16.8% (21.9%) | 18.9% (28.2%) |
| Median crystalline-state loss (dB cm) | below floor | 5.6 |
| Largest computed loss, any state (dB cm) | 51 | |
| Dies trimmed inside ps km, linear model | 99.66% | 93.3% |
| Median added loss after trimming (dB cm) | 3.1 | |
| Bit patterns used, of 16 | 7 | 16 |
Table 1. The two claddings over the converged sweep and the linear Monte Carlo of Sec. VI-A (four-segment set on the mask). Losses are modal propagation losses computed with the optical constants of Ref. [17] ( for SbS).
Figure S6 gives the distributions of , , mode area and loss.
Switching contrast, loss and the choice of material
Crystallising the strip raises the cladding index and pulls the mode upward, changing both and its curvature. For a representative SbS geometry (, , , nm), at 1550 nm moves from ps km amorphous to ps km fully crystalline, passing 135, 459 and 1026 ps km at , 0.5 and 0.75 (Figure 3). Because the response is strongly nonlinear in , the device uses only the two end states.

Figure 3. Dispersion switching of one Sb₂S₃-clad waveguide (, , , nm). (a) at five crystalline fractions. (b) at 1550 nm against . Shaded: anomalous dispersion.
Switching contrast scales sharply with strip thickness (Figure 4(a)). The median for SbS rises from 7.6 ps km at nm to 181 at 40 nm, 1863 at 80 nm and 1957 ps km at 120 nm; SbSe reaches similar values at roughly half the thickness, consistent with its larger index contrast.

Figure 4. Switching range and its cost. (a) Median against strip thickness; bands: interquartile range. (b) Switching contrast against crystalline-state loss, bubble area ; Sb₂S₃ losses ( in the adopted data) sit at the numerical floor (shaded). (c) Fraction of as-deposited geometries with an in-band zero-dispersion wavelength, against film thickness.
The materials separate on loss (Figure 4(b)). With the extinction coefficient of the adopted dataset, zero for SbS in both states [17], the computed modal loss of the SbS sweep never exceeds dB cm, below the level at which the pole fit is meaningful; crystalline SbSe has a median of 5.6 dB cm and a maximum of 51 dB cm. The SbS value reflects the optical constants used, not a lossless film: residual absorption of deposited films and edge scattering are not modelled, and junction loss is treated in Sec. IV-C. Within these constants SbS provides the range within a loss budget compatible with a resonator or a long nonlinear interaction (Table I), and the rest of the paper uses it.
The sign of the swing depends on the strip thickness. At nm, crystallising SbS moves in the normal direction, so a device biased anomalous can be trimmed toward zero. At nm the mean swing over the sensitivity set is ps km, in the anomalous direction, possibly because the thin crystalline layer changes the curvature of the mode more than its confinement. The segment library stays in the thick-strip regime, where the sign is consistent.
Segmented Binary Control
Dispersion of a segmented guide
Consider a waveguide of total length divided into segments of length , each with its own independently switched strip. If the junctions are adiabatic (checked in Sec. IV-C), the accumulated phase is
where is the propagation constant of segment in its current state. Every -derivative inherits the linearity, so the device dispersion of order is
The device therefore has discrete dispersion states, each addressed by an -bit pattern. A single strip switched continuously in traces a curve in the plane, so fixing fixes ; segments whose switching vectors point in different directions span a region with area.
The segment library and the reachable set
A library of eleven SbS segment types was solved on the nominal waveguide ( nm, nm), covering three strip widths and four strip thicknesses in both states. Each entry is a tuple per state, and any candidate device follows from (3) without further mode solving.
Figure 5 compares the process sensitivities with the authority of one segment, over 28 nominal geometries at the strip slice nm, nm. The derivatives and are properties of the ridge and vary slowly with the strip, so this slice sets the error budget.

Figure 5. Fabrication sensitivity and trim authority. (a) and (b) against film thickness, per material. (c) Nanometres of each error that one segment absorbs; dashed lines are the process windows (15 nm in width, 8 nm in thickness).
Enumerating the 16 bit patterns of the four segments selected in Sec. VI-A gives the point cloud of Figure 6(a). The set spans to 580 ps km in and 0.39 to 2.51 ps km in , and the widest interval at a single is 0.44 ps km. The area comes from segments moving the device in different directions through the plane (Figure 6(b)).

Figure 6. (a) The 16 states of the four-segment trimmer in the plane, coloured by the number of crystallised segments, with their convex hull. (b) Per-segment switching vectors from amorphous (open) to crystalline (filled). (c) interval reachable at each ; the bar marks the widest, 0.44 ps km at ps km.
A full-wave check of the segment junctions
Equations (2) and (3) neglect reflection, mode conversion and radiation where the strip changes. This was tested with eigenmode expansion [26] in Tidy3D [24], with abrupt junctions and each cross-section described by its three lowest modes, sorted by TE fraction and solved at 11 frequencies between 1380 and 1720 nm; the fundamental mode of every cross-section matches the mode solves to six decimal places in . Each segment is 100 m long between 20 m bare leads. The part of the transmitted phase not explained by the length-weighted model was fitted with a quartic in frequency,
and its curvature is reported as a contribution to the of a device of four 500 m segments.
The junctions add at most 0.36 ps km to for any single segment and 1.24 ps km for the four-segment device in pattern 1111; for patterns 0000, 1010 and 1100 the contribution is 0.12, 0.02 and 0.003 ps km [Fig. 7(b)]. All are an order of magnitude below the numerical floor, so within that floor Eq. (3) holds for the 2 mm device.
A single segment between bare leads transmits TE with a loss of 0.06 to 0.65 dB amorphous and 0.16 to 2.45 dB crystalline [Fig. 7(a)]. The largest loss belongs to the crystalline 900/80 nm strip, whose fundamental index, 2.03, differs most from that of the bare ridge, 1.79. Reflection stays below dB in every run, so the lost power is coupled to higher modes or radiated. Inside the device the index steps between neighbouring strips are smaller, and the four segments together lose 0.45 dB (0000) to 1.72 dB (1111), a third to a half of the sum for isolated segments [Fig. 7(c)]. With three modes per cross-section these values indicate the scale of the junction loss rather than converged figures. Tapers at the strongest segment, or the loss-weighted library of Sec. VI-G, would reduce it without changing Eq. (3).
A Physics-Informed Surrogate for the Design Space
The segment states do not move rigidly with the die error (Sec. VI-C), so choosing a pattern for a measured die needs the library at the die’s own geometry, and inferring the die’s from a test-ring group index needs and . A regression that fits values well can still have derivatives of the wrong sign. The surrogate below supplies both, with derivatives constrained by relations that hold exactly for the Maxwell problem solved here.
Exact relations for this waveguide
The relations assume a non-magnetic, -invariant guide whose media have diagonal permittivity tensors aligned with the guide axes; those built on Eq. (4) also assume a lossless guide. Both hold exactly for the training data: the LiNbO Sellmeier model, the constant-index silica and the adopted SbS constants () are lossless. The scale invariance used in Eq. (6) holds with loss as well, but its expression through the confinement factors does not, so a lossy cladding such as crystallized SbSe would need the complex generalisation; the surrogate is trained on SbS only. The first-order change of the propagation constant caused by a permittivity change at fixed is [27, 28]
so a uniform index change of region (or tensor component) gives the confinement factor [29, 30, 31]
Maxwell’s equations contain no length scale: the curl–curl equation depends on the indices only through and on the four lengths only through [32]. Euler’s theorem then gives, at fixed material indices,
and adding the material dispersion gives , the equality of group and energy velocity [33], with the bulk group index of medium . Subtracting the two removes the non-dispersive media (constant-index silica and air) and leaves a first-order partial differential equation linking the frequency derivative of to its gradient in the geometry,
which we call the dispersion–geometry identity. Its left side is constrained by the spectra and its first right-hand term by the geometry sweep. The LiNbO material terms are small ( for and for at 1550 nm), so an error of a few per cent in moves the right-hand side by less than . For the strip, Eqs. (1) and (5) give

Figure 7. Eigenmode-expansion check of the segment junctions at 1550 nm. (a) TE insertion loss of each segment alone, amorphous (open) and crystalline (filled). (b) Junction contribution to the of the 2 mm device against the 10 ps km floor. (c) Loss of the four-segment device against the sum of its segments, per bit pattern.
with and , so rises monotonically with . Finally, perturbation theory for shifting material boundaries [34] gives a surface integrand that is positive when a higher-index medium advances into a lower-index one. This fixes the sign of , and (given ) but not of , because raising the LiNbO top face puts LiNbO where SbS was beneath the strip.
The discretised solver obeys these identities closely (Table II), so imposing them does not conflict with the data. was taken by central differences on a frozen grid, and by scaling geometry, domain, grid and absorbing layers together, which keeps the discrete problem exactly scale-invariant. Confinement factors from the field integrals agree with the finite-difference values to 0.1% for and , and to 1.6–4.5% for , whose field integral over a thin, high-index strip depends on how the mesh resolves the field discontinuity at its faces; the network therefore takes from Eq. (8) applied to itself. On the grid sweep, central differences (55 interior points) leave a LiNbO term in Eq. (7) of , against about expected from and .
Model and training
One multilayer perceptron maps to , and the five coefficients of , a quartic in ; , , and follow by exact integration and by exact differentiation, so and are consistent by construction. The output is written as with , which makes the strip-free limit exact and state-independent. A second head returns . Input derivatives are exact, by forward-mode automatic differentiation; because the spectra are linear in the predicted coefficients, one Jacobian–vector product along returns the whole spectrum of .
| Identity | amorphous | crystalline |
| Eq. (6), scale invariance | ||
| , energy velocity | ||
| Eq. (7), finite-difference | ||
| Eq. (7), field-integral |
Table 2. Closure of the exact identities on the mode solver (production mesh) at the nominal segment (, , nm, 1550 nm), in units of index.
The dispersion calculus, the strip-free limit and are built into the architecture, as in hard-constrained networks [35]. The residual of Eq. (7), with from Eq. (8), and the sign conditions are penalised on unlabelled collocation points over the design box, as in physics-informed learning [36, 37, 38]; fitting derivatives as well as values is a form of Sobolev training [39]. Earlier physics-aware photonic regressors embedded the effective-index approximation [40]; every constraint used here is exact within the assumptions above. Each residual is divided by a tolerance near the solver noise, for example 10 ps km on and on Eq. (7); the full list is in Sec. S6.
Training uses the SbS rows of the grid sweep (1546) and the space-filling rows of the second dataset (496); the jitter set is held out entirely as truth for Sec. VI-C. The test split is grouped by geometry. Each member of a five-member ensemble [41] is trained in three stages, identical for all variants compared below: 20 000 full-batch Adam epochs [42] on the data, 3000 epochs with the variant’s own loss, and an L-BFGS polish [43]. Baselines are a Gaussian process with a Matérn-5/2 kernel [44], gradient-boosted trees [45], an unconstrained network that predicts the grid and as separate outputs, and the same spectral network trained without physics terms.
Accuracy and physical consistency
The models are compared on interpolation inside the training domain, physical consistency, and extrapolation outside it (Table III). No single model is best on all three.
| Model | Interp. MAE | Consistency gap | Sign (%) | Residual | Extrap. MAE |
| PI | 4.8 | 0 | 1.1 | 8.3 | |
| No physics | 4.4 | 0 | 6.3 | 13.0 | |
| GP | 3.4 | – | n/a | n/a | 22.2 |
| MLP | 2.1 | 0.021 | n/a | n/a | – |
| Trees | 38 | – | n/a | n/a | – |
| ANN (non-physics) | 5.2 | n/a | n/a | n/a | 36.4 |
Table III. The surrogates on the three criteria. Interpolation: mean absolute error (MAE, ps km) of at 1550 nm on the 405 held-out rows. Consistency: mean gap between the predicted and the derivative of the predicted (ps km), and, at random design points, the fraction of sign conditions violated and the RMS residual of Eq. (7). Extrapolation: MAE at nm after training on nm. PI, physics-informed; GP, Gaussian process; MLP, unconstrained network; n/a, not defined for the model; –, not computed. Network predicting the spectrum without physics terms, trained on both claddings with a separate grouped split (606 held-out rows; 756 rows at nm). Its early stopping used validation rows that included nm, so its extrapolation error is likely an underestimate.
a) Interpolation: On the 405 held-out rows the physics-informed ensemble predicts at 1550 nm with a mean absolute error of 4.8 ps km, with 90% of its predictions inside the ps km floor (Figure 8(a,Figure 8c)). The same network without physics terms reaches 4.4 ps km, the Gaussian process 3.4, the unconstrained network 2.1 and gradient-boosted trees 38. All but the trees are inside the floor, so the physics terms bring no gain in accuracy here; the unconstrained network is the most accurate, at the price of a that is not the derivative of its own . A variant with only the guidance bounds (, ) coincides with the network without physics terms, because the data already satisfy the bounds.

Figure 8. The physics-informed surrogate. (a) Predicted against solved on 405 held-out rows (asinh scale); shaded: ps km. (b) Six criteria, with and without physics terms, each normalised so that 1 marks the better model. (c) Absolute error on held-out rows. (d) Relative error of the geometric derivatives against the 65 jitter dies. (e) Waveguide and material parts of for the 600/40 nm segment, amorphous (solid) and crystalline (dashed). (f) Solved device of every jitter die for four ways of choosing the pattern; red: outside ps km.
b) Consistency: Without physics terms the network violates the sign conditions at 6.3% of random design points and leaves an rms residual of in Eq. (7), with 1.5% of points outside its tolerance; with them the figures are 1.1%, and none (Figure 8(b)). These checks need the predicted spectrum, its geometric gradient and the confinement factors, so they apply to the spectral networks only. The unconstrained network’s differs from the derivative of its own by 0.021 ps km on average, comparable to its error of 0.018 ps km.
c) Extrapolation: Trained only on nm and asked for nm, the physics-informed ensemble errs by 8.3 ps km, the network without physics terms by 13.0 and the Gaussian process by 22.2 (Figure 8(b)); the unconstrained network was not run on this split. A separately trained spectral network without physics terms, fitted to both claddings, gives 5.2 ps km in interpolation but 36.4 ps km at nm (Table III). This is one held-out slice in one direction, so the result is indicative.
d) Derivatives and calibration: Against the derivatives fitted to the 65 jitter dies, the median relative error of is 2.0% with the physics terms and 2.3% without, that of 1.2% for both, and both width derivatives are reproduced to about 1% (Figure 8(d)). The raw ensemble spread is narrower than the error: its 90% interval covers 69% of the held-out values (normalised errors with a standard deviation of 2.9), against 92% for the Gaussian process, and at nm 76%, against 64% without the physics terms. The pattern choice below therefore uses the spread inflated by a factor of 1.93, calibrated for 90% coverage on the held-out rows.
Differentiating Eq. (7) in splits exactly into a waveguide part and a material part, which the surrogate evaluates from its own gradient (Figure 8(e)). For the 600/40 nm segment at 1550 nm the amorphous of ps km (Tidy3D: ) is waveguide and material. On crystallisation both parts rise by 90 ps km, so half of the switching contrast is the dispersion of crystalline SbS itself and half is the change of the mode.
Die-Specific Inference and Trimming
Selecting the segment set
The four of the eleven segment types placed on the mask are chosen once, before fabrication. All 330 four-element combinations were ranked by the fraction of Monte Carlo dies (below) brought inside specification. The best set, , 600/40, 900/40 and 900/80 nm, is tied by the second-best and followed closely by the next, so several sets perform similarly (Sec. S4 and Fig. S7 of the supplementary material).
A linear Monte Carlo
Each of 5000 synthetic dies per material draws a ridge-width error and a film-thickness error from normal distributions with nm and nm, representing electron-beam linewidth control and commercial wafer thickness uniformity. The as-fabricated is the nominal value plus a first-order expansion in the two errors with the sensitivities of Figure 5, and the trimmer picks the bit pattern that brings closest to zero. A die for which no pattern lands inside ps km is untrimmable.
For SbS, 99.66% of dies are trimmed into specification and the standard deviation of falls from 24.2 to 11.1 ps km, at a median added loss of dB cm (Figure 9). The residual spread is set by the discreteness of the 16 states (see Sec. VI-G). The nominal design is biased anomalous, at ps km, because crystallisation moves only in the normal direction at these strip thicknesses; the trim removes this designed offset and the process spread together. The untrimmable 0.34% of SbS dies occupy one corner of the plane (Figure 9(b)), a region that could be screened at wafer test. Seven of the 16 patterns are used. SbSe uses all 16 and leaves 6.7% of dies untrimmable, at a median loss of 3.1 dB cm. The neglected quadratic terms have a median of 0.4 to 0.6 ps km and reach 11.3 ps km at three standard deviations (Sec. S5), so the linear model holds for the bare ridge over most of the distribution; whether the segment states follow it is tested next.

Figure 9. Monte Carlo over 5000 dies per material; top , bottom . (a,d) before (grey) and after (colour) trimming; shaded: . (b,e) Error plane with and process contours; untrimmable dies in dark red. (c,f) Dies per bit pattern.
A full-wave check at fabricated geometries
The Monte Carlo treats a die error as a rigid shift: a die is assumed to move every pattern of the nominal library by as much as it moves the bare ridge. Each of the 65 jitter dies has its own solved library, so this can be tested directly.
The segment states do not move together [Figure 10(a)]. The amorphous segments follow the bare ridge, with between and ps km per nanometre against for the bare ridge, but the crystalline 900/80 nm segment has a slope of ps km per nanometre and a width sensitivity of ps km per nanometre, against 0.4 to 0.7 for the others, and its switching contrast ranges from 1937 to 2096 ps km over the box.

Figure 10. Mode solves at 65 die geometries in the box. (a) Change of each segment state’s from the nominal die against the film-thickness error; filled: crystalline. (b) Device for four ways of choosing the pattern; red: out of specification. (c) Error plane for the rigid-shift rule with , and contours. (d) Device predicted by the rigid-shift rule against the solved value.
Four ways of choosing the pattern were scored on the solved device of each die, Eq. (3) with equal lengths [Figure 10(b) and Table IV]. With no pattern written no die is in specification, as designed, and the single pattern best on the nominal die, 1100, helps only dies near the nominal. The rigid-shift rule, given each die’s error exactly, brings 83% of the box into specification (94% weighted by the process density, with an effective sample size of 22 dies). It mispredicts the device by a standard deviation of 12.2 ps km and by up to 38 ps km [Figure 10(d)], and all eleven of its failures lie
| Pattern choice | In spec (%) | SD of (ps km) | ||
| box | weighted | box | weighted | |
| None (0000) | 0 | 0 | 41.2 | 22.6 |
| Fixed, nominal best (1100) | 34 | 63 | 46.7 | 25.7 |
| Rigid shift of nominal library | 83 | 94 | 17.1 | 13.7 |
| Library of the die (oracle) | 100 | 100 | 11.8 | 11.4 |
Table IV. Device at 1550 nm on the 65 solved jitter dies for four ways of choosing the bit pattern. In spec: ps km; weighted by the process density.
more than from the nominal die [Figure 10(c)]. Choosing from the die’s own library brings all 65 dies in, with a residual standard deviation of 11.8 ps km.
Inferring the die error and choosing the pattern
The die error is read from test rings whose free spectral range gives the group index at 1500, 1550 and 1600 nm, with a resolution of . Each jitter die’s true plus measurement noise is inverted on the surrogate, the pattern with the highest probability of landing within ps km is chosen, and the outcome is scored on the Tidy3D device of that die [Figure 8(f)]. With one bare ring the surrogate brings 91% of the dies (59 of 65; 88% weighted by the process density) into specification in a single noise draw, against 88% (87% weighted) when the Gaussian process chooses from the
| Rings | sd() (nm) | sd() (nm) | Corr. | In spec (%) | Weighted (%) |
| bare | 4.4 | 8.1 | 0.98 | 88.8 | 93.0 |
| bare + 300/80 a | 1.17 | 1.37 | 0.75 | 88.8 | 88.1 |
| bare + 300/80 c | 0.66 | 0.81 | 0.01 | 92.9 | 91.8 |
| bare + 600/40 a | 1.08 | 1.27 | 0.68 | 90.9 | 90.8 |
| bare + 600/40 c | 0.66 | 0.85 | 0.04 | 94.5 | 92.8 |
| bare + 900/40 a | 0.96 | 1.17 | 0.57 | 92.9 | 90.6 |
| bare + 900/40 c | 0.58 | 0.90 | −0.07 | 96.0 | 96.4 |
| bare + 900/80 a | 0.52 | 0.89 | −0.23 | 96.2 | 97.0 |
| bare + 900/80 c | 0.41 | 0.91 | −0.53 | 96.8 | 99.1 |
| true geometry | – | – | – | 100 | 100 |
Table V. DIE ERROR FROM THE TEST RINGS AT THE NOMINAL DIE, AND THE LOOP SCORED ON THE TIDY3D DEVICE OF THE 65 JITTER DIES (TEN NOISE DRAWS EACH). IN SPEC: PS KM; A, AMORPHOUS; C, CRYSTALLINE STRIP.
same inferred geometry. Given the true geometry the surrogate brings every die in, against 83% (94% weighted) for the rigid-shift rule with the same information, so the per-die model removes the error of the rigid shift and what remains is the inference. One ring recovers the geometry with rms errors of 4.8 nm in and 10.0 nm in , and three of the six failures miss the band by less than 1 ps km.
A second test ring
One bare ring constrains mainly one combination of the two errors: a linearised Fisher analysis with the process prior gives posterior standard deviations of 3.3 nm in and 6.0 nm in with a correlation of 0.99, because the group index moves with and in nearly the same ratio at all three wavelengths. A second bare ring 300 nm wider helps only partly (1.3 and 2.7 nm, correlation 0.97). A ring carrying a phase-change strip needs no new structure, since the mask already holds eight strip states, and its group index at every jitter die is in the jitter solves, so the loop is scored entirely on Tidy3D. A crystalline strip draws the mode towards the cladding and changes the ratio of the two sensitivities. Table V lists every pairing, with the surrogate’s own group-index error (2.9 × rms on the held-out rows) added to the ring resolution as noise. Every crystalline strip ring removes most of the correlation, and the crystalline 900/80 nm ring gives the smallest error, 0.41 nm in and 0.91 nm in [Fig. 11(a)].
In the loop each ring is read at the three wavelengths, the die error is taken from the posterior on a grid over , and the pattern is chosen from 48 posterior samples; each die is measured ten times with independent noise. One bare ring brings 88.8% of the trials within ps km (93.0% weighted). Adding the crystalline 900/80 nm ring raises this to 96.8% (99.1% weighted), with 3.2% false successes, and the rms errors of the recovered geometry fall from 6.1 and 11.7 nm to 1.7 and 4.2 nm [Fig. 11(b)]. The pair is insensitive to the ring resolution: at 1 × and 5 × it gives 96.9% and 96.6%, while one bare ring falls from 91.5% to 79.2%. The remaining failures therefore appear to be set mainly by the surrogate’s own group-index error rather than by the measurement.
Sensitivity to a systematic solver offset
The success rates above are scored on the solver’s own and so inherit the numerical floor of Sec. II-B. A systematic error shifts every die the same way. To bound its effect, a common offset was added to the solved device of every die after the pattern had been chosen, modelling a fabricated device that differs from the solver by while the surrogate, trained on the solver, does not (Table VI). With the two rings the success rate is 96.8% at , 99.8% at +5 and 98.6% at +10 ps km, and 86.0% and 73.5% at −5 and −10 ps km. The asymmetry follows from the trimmed distribution, with mean −4.8 ps km and standard deviation 11.6 ps km: a positive offset recentres it, and a negative one pushes its lower tail out of the band. Tightening the band has the same effect: the two rings bring 86.0% of the trials within ps km and 72.2% within ps km, against 78.8% and 60.6% for one bare ring. An offset could be measured on a first fabricated die and removed by shifting the target of the pattern choice.
Designing the segment library with exact derivatives
Because the surrogate is differentiable in the strip geometry, the segment library itself can be designed. A library is four strips (, ) inside the training box and four length fractions of the 2 mm device. It is scored on 1024 dies drawn from the process distribution, with the pattern chosen by the calibrated ensemble and each ensemble member taken in turn as the truth. A smooth version of this yield is maximised with Adam (four starts of 150 steps per loss weight) at ps km, where designs differ; at ps km every library considered, including the current mask, brings more than 99.5% of the dies in. Junction loss enters through a proxy, over the junctions of a pattern, with dB fitted to the single-segment eigenmode-expansion runs of Sec. IV-C; it reproduces those runs (0.06 to 2.45 dB) and the four-segment runs (0.45 to 1.72 dB) to within 0.19 dB. On the current mask the yield estimate matches the Tidy3D dies at ps km (100% for both) and lies 3 to 5 points above them otherwise (58.5% against 53.8% at ps km, and 99.7% against 96.8% with the two rings).
At ps km the current mask brings 58.9% of the dies in with the true geometry (58.1% with two rings), because its 16 levels cluster [Fig. 11(d)]. Keeping its strips but making the lengths binary (0.13, 0.53, 0.27 and 0.07 of the device for the 300/80, 600/40, 900/40 and 900/80 nm strips) spreads the levels and gives 84.0% (83.1%). Optimising strips and lengths together gives 99.8% (99.1%). With the junction loss weighted at 0.1 per dB, the optimiser replaces the 900/80 nm strip, whose crystalline step dominates the loss, by a 448/64 nm strip; the library 300/74, 543/40, 899/42 and 448/64 nm with lengths 0.13, 0.50, 0.23 and 0.14 keeps 99.8% (99.0%) while the worst pattern loses 0.28 dB instead of 2.13 dB [Fig. 11(c)]. Its patterns span to 116 , narrower than the to 576 of the current mask but more than 100 on either side of zero. At 1.5 times the process spread it keeps 97.1% (96.0%), against 59.7% for the current mask. In the model, an offset of up to the 10 floor would keep dies placed within inside . The optimised strips lie inside the training box; a mode solve of the new strips and an eigenmode-expansion run with the new lengths are the next checks.
| Pattern choice | Truth | Trials | In spec (%) | w (%) |
| Rigid shift, exact error | LM | 5000 | 99.66 | – |
| Rigid shift, exact error | T3D | 65 | 83 | 94 |
| Library of the die (oracle) | T3D | 65 | 100 | 100 |
| Surrogate, exact geometry | T3D | 65 | 100 | 100 |
| Surrogate, bare ring | T3D | 650 | 88.8 | 93.0 |
| Surrogate, bare + 900/80 c | T3D | 650 | 96.8 | 99.1 |
| with | T3D | 650 | 73.5 | 73.4 |
| with | T3D | 650 | 98.6 | 98.7 |
| Optimised library, two rings, | ens. | 1024 | 99.0 | – |
Table VII. Success rates of the pattern-choice rules. Band except where stated; W, weighted by the process density; LM, linear model; T3D, Tidy3D device of the jitter dies; ens., surrogate ensemble as truth. Trials: dies, times ten noise draws for the ring rows and five ensemble members for the last row.
Summary of success rates
Table VII collects the success rates, which differ in what they are scored on. The linear Monte Carlo is scored on its own model; on full-wave solves the same rigid-shift rule reaches 94% of the weighted dies rather than 99.66%. The jitter box over-represents the tails of the process distribution (effective sample size 22 dies), which is why the weighted values are higher. The library-design row is scored on the surrogate ensemble, at a tighter band, and awaits a full-wave check.
Thermal Feasibility of Laser Writing
The trim is applied once, after the die has been measured, so it needs no integrated heater: a segment could be crystallised by a focused beam, as done for SbS films with a 532 nm laser, where continuous-wave exposure recrystallised the film within 7 to 600 ms at 1 to 30 mW [46]. Only the amorphous-to-crystalline direction is needed, so no melt-quench step is involved. The estimate below is thermal only: it determines the absorbed fraction, whether the strip can be held above its crystallisation temperature without melting, and whether the neighbouring strip stays below its own. Crystallisation kinetics and the uniformity and repeatability of a scanned write are not modelled.
a) Absorption and heating. A three-dimensional FDTD solve at 532 nm with a Gaussian beam of 0.6 waist centred on one 600/80 nm strip gives an absorbed fraction of 0.237 with the strip amorphous and 0.404 crystalline, and in the neighbouring strip 1.5 away [Fig. 12(a)]. Transient heat conduction was solved in the plane through the two strip axes; a three-dimensional transient solve for one case gives a peak rise 0.77 times that of the slice [Fig. 12(b)], and powers are quoted both from the slice and divided by that ratio. The model, its constants and a thin-film check of the absorption are given in Sec. S7 and Table S4.
b) A predicted write window: With constant material properties the temperature rise is linear in absorbed power, and three conditions fix the window: the amorphous strip must reach the crystallisation temperature, 280 °C [46]; the strip must stay below the melting point, 550 °C [47], after it has crystallised and absorbs times more of the beam; and the neighbour must stay at least 50 K below the crystallisation temperature. The second condition binds and limits the window to a factor in power, independent of dwell time, so for a fixed power would have to be held to about . For a 1 ms dwell, the regime of Ref. [46], the predicted window is 1.9 to 2.3 mW of incident power from the slice and an estimated 2.5 to 3.0 mW after the three-dimensional correction [Fig. 12(c,d)], which rests on a single three-dimensional case. At the top of the window the neighbouring strip heats by 21, 13 and 6 K for gaps of 1, 2 and 4 . Shorter dwells need 4.6 to 5.5 mW at 1 and 11.5 to 13.8 mW at 100 ns after correction; there the crystallisation kinetics of the film [48] also enter. The window assumes constant material properties and complete crystallisation once the threshold is reached; both, and the effect of the writing conditions on the crystalline index, need measurement.
Discussion
The largest uncertainty in the two-ring success rate is the numerical floor. If the solver error is systematic, the rate lies between 73.5% and 99.8% for offsets within (Table VI); a calibration on a first fabricated die, or a library designed for a tighter band (Sec. VI-G), would recover most of the margin. The SbS optical constants also vary between laboratories [23], which shifts the absolute dispersion values, and the negligible material loss holds for the adopted constants () rather than for any deposited film. In the present model the dominant optical cost is the insertion loss of the abrupt junctions, mostly at the crystalline 900/80 nm segment; a converged figure needs a larger model basis or a finite-difference time-domain calculation.
| (ps km) | −10 | −5 | 0 | +5 | +10 |
| Bare + 900/80 c | 73.5 | 86.0 | 96.8 | 99.8 | 98.6 |
| w | 73.4 | 87.6 | 99.1 | 99.9 | 98.7 |
| Bare ring only | 70.0 | 84.8 | 88.8 | 87.7 | 84.9 |
| w | 73.1 | 88.1 | 93.0 | 90.9 | 91.1 |
Table VI. SUCCESS RATE WITHIN PS KM (%) ON THE 65 JITTER DIES, TEN NOISE DRAWS EACH, WITH A COMMON OFFSET ADDED TO THE SOLVED DEVICE , w. WEIGHTED BY THE PROCESS DENSITY.
The second test ring is the main practical element of the inference. Film and width errors are nearly degenerate in the group index of a bare ring, and a ring carrying a crystalline strip, already on the mask, removes most of the degeneracy. With two rings the remaining failures follow the surrogate’s group-index error, so more training data near the nominal die is the most direct way to reduce them. Inside the training domain the physics terms do not improve accuracy; they add

Figure 11. Test rings and library design. (a) Posterior ellipses of the die error: process prior (dashed), one bare ring, and bare plus the crystalline 900/80 nm ring (thin: the other strip rings). (b) Solved device of the jitter dies after the pattern choice with one ring, two rings and the true geometry; shaded: ps km. (c) Model yield at ps km against worst-pattern junction loss; bubble area: range. (d) The 16 levels at the nominal die for equal, binary and optimised lengths; shaded: ps km.

Figure 12. Thermal model of a 532 nm write. (a) Absorbed fraction in the target strip, FDTD and thin-film estimate. (b) Peak temperature rise for 3 mW absorbed, three- and two-dimensional models. (c) Peak temperature against incident power for a 1 ms dwell: target strip amorphous (solid) and crystallised (dotted), neighbour at 1, 2 and 4 m (dashed); shaded: predicted write window. (d) Predicted window against dwell time.
consistency between and , derivatives with the correct signs, and a smaller error in the one extrapolation tested. The identities are exact for the lossless, non-magnetic guide solved here and would need their complex form for an absorbing cladding.
The sweep is of a fully etched ridge, whereas most thin-film lithium niobate processes leave a slab; a partial etch changes the confinement and therefore , and the numbers here apply to the geometry as swept. The write step is modelled only through absorption and heat flow. The time at temperature needed for full crystallisation, the uniformity along a 500 segment written by a scanned spot, the repeatability from segment to segment and any change of the crystalline index with the writing conditions are informed by work on SbS kinetics [48], laser writing [46] and back-end-of-line integration [21], but are not predicted here. Because the trim is one-way, cycling endurance does not enter. All results are numerical.
Conclusion
In simulation, a binary SbS cladding divided into a few independently switched segments moves the group-velocity dispersion of a thin-film lithium niobate waveguide over a range large compared with the error of a commercial wafer, with two degrees of freedom, so and need not be traded against each other. Four segments reach to ps km in , and eigenmode expansion indicates that the junctions leave the length-weighted model intact to within ps km.
Because the segment states move with the die error, the bit pattern is best chosen from a model of the die. A spectral surrogate constrained by exact identities of the lossless Maxwell problem provides that model with consistent and , geometric derivatives within about 2% of the solver, and an error of 8.3 rather than 13.0 ps km at an unseen film thickness. Read through a bare ring and a ring carrying the crystalline 900/80 nm strip, it brings 96.8% of the dies (99.1% weighted) within ps km on full-wave truth, and every die when the geometry is known; the solver floor, rather than the inference, sets the main uncertainty.
The same derivatives make the segment library a design variable: jointly optimised strips and lengths bring 99% of the model dies within ps km while the worst-case junction loss falls from 2.1 to 0.3 dB. A thermal model predicts a candidate write window of 2.5 to 3.0 mW at 532 nm for a 1 ms dwell; the kinetics, uniformity and repeatability of the write remain to be measured. A fabricated trimmer with its two test rings is the next step, and the same procedure could address other dispersion targets, such as a comb pump or a frequency-conversion bandwidth.
Author Contributions
S.D.: conceptualization, methodology, software, formal analysis, writing of the original draft. A.K.M.: machine-learning modelling, writing (review and editing), manuscript preparation. S.S.: supervision.
Disclosures
The authors declare no conflicts of interest.
Data Availability
The simulation results and the scripts that produce every figure and table are available from the authors on reasonable request.
Acknowledgments
References
- [1]M. Zhang, C. Wang, R. Cheng, A. Shams-Ansari, and M. Lončar, “Monolithic ultra-high-q lithium niobate microring resonator,” Optica, vol. 4, no. 12, pp. 1536–1537, 2017.
- [2]M. Zhang, M. Zhang, X. Chen, M. Bertrand, A. Shams-Ansari, S. Chandrasekhar, P. Winzer, and M. Lončar, “Integrated lithium niobate electro-optic modulators operating at cmos-compatible voltages,” Nature, vol. 562, no. 7725, pp. 101–104, 2018.
- [3]D. Zhu, L. Shao, M. Yu, R. Cheng, B. Desiatov, C. Xin, Y. Hu, J. Holzgraf, S. Ghosh, A. Shams-Ansari et al., “Integrated photonics on thin-film lithium niobate,” Advances in optics and photonics, vol. 13, no. 2, pp. 242–352, 2021.
- [4]A. Boes, L. Chang, C. Langrock, M. Yu, M. Zhang, Q. Lin, M. Lončar, M. Fejer, J. Bowers, and A. Mitchell, “Lithium niobate photonics: Unlocking the electromagnetic spectrum,” Science, vol. 379, no. 6627, p. eabj4396, 2023.DOI
- [5]Y. Song, Y. Hu, X. Zhu, K. Yang, and M. Lončar, “Octave-spanning kerr soliton frequency combs in dispersion-and dissipation-engineered lithium niobate microresonators,” Light: Science & Applications, vol. 13, no. 1, p. 225, 2024.DOI
- [6]Z. Gong, X. Liu, Y. Xu, and H. X. Tang, “Near-octave lithium niobate soliton microcomb,” Optica, vol. 7, no. 10, pp. 1275–1278, 2020.
- [7]Y. He, Q.-F. Yang, J. Ling, R. Luo, H. Liang, M. Li, B. Shen, H. Wang, K. Vahala, and Q. Lin, “Self-starting bichromatic linbo3 soliton microcomb,” Optica, vol. 6, no. 9, pp. 1138–1144, 2019.
- [8]M. Yu, B. Desiatov, Y. Okawachi, A. L. Gaeta, and M. Lončar, “Coherent two-octave-spanning supercontinuum generation in lithium niobate waveguides,” Optics letters, vol. 44, no. 5, pp. 1222–1225, 2019.DOI
- [9]J. Mishra, M. Jankowski, A. Y. Hwang, H. S. Stokowski, T. P. McKenna, C. Langrock, E. Ng, D. Heydari, H. Mabuchi, A. H. Safavi-Naeini et al., “Ultra-broadband mid-infrared generation in dispersion-engineered thin-film lithium niobate,” Optics Express, vol. 30, no. 18, pp. 32 752–32 760, 2022.DOI
- [10]M. Jankowski, C. Langrock, B. Desiatov, A. Marandi, C. Wang, M. Zhang, C. R. Phillips, M. Lončar, and M. M. Fejer, “Ultrabroadband nonlinear optics in nanophotonic periodically poled lithium niobate waveguides,” Optica, vol. 7, no. 1, pp. 40–46, 2020.
- [11]O. Hefti, J.-E. Tremblay, A. Volpini, Y. Koyaz, I. Prieto, O. Dubochet, M. Despont, H. Zarebideki, C. Caer, J. Berney et al., “Fabrication-tolerant modal phase matching for all-optical frequency conversion in layer-poled thin film lithium niobate waveguide,” APL Photonics, vol. 10, no. 10, 2025.
- [12]Y. Wu, H. Sun, Y. Yu, B. Xiong, W. Ma, J. Zhang, Z. Zheng, and T. Chu, “Lossless non-volatile post-fabrication trimming of silicon photonics through on-chip high-temperature annealing of undercut waveguides,” Laser & Photonics Reviews, vol. 20, no. 6, p. e01982, 2026.DOI
- [13]Y. Wu, H. Sun, B. Xiong, Y. Y. Zhang, Z. Zheng, W. Ma, and T. Chu, “Lossless, non-volatile post-fabrication trimming of pics via on-chip high-temperature annealing of undercut waveguides,” 2025. [Online]. Available: https://arxiv.org/abs/2506.18633
- [14]Y.-D. Wang, Z.-N. Tian, Y.-C. Li, Z.-D. Zhang, L.-C. Wang, and Q.-D. Chen, “Phase customization in photonic integrated circuits with trimmed waveguides,” Optics Letters, vol. 47, no. 22, pp. 5889–5892, 2022.DOI
- [15]C. Lian, Y.-S. Huang, H. Sun, H. Yu, C.-C. Popescu, T. Gu, S. A. Vitale, I. Takeuchi, J. Hu, and C. Ríos, “Phase and amplitude trimming of photonic integrated circuits using phase change materials,” in 2022 IEEE Photonics Society Summer Topicals Meeting Series (SUM). IEEE, 2022, pp. 1–2.
- [16]J. Pang, R. Chen, S. Gao, S. Zhang, and T. Cao, “Non-volatile sb2s3–ito post-fabrication trimming for high-precision optical convolutional neural networks,” Laser & Photonics Reviews, p. e71146, 2026.
- [17]M. Delaney, I. Zeimpekis, D. Lawson, D. W. Hewak, and O. L. Muskens, “A new family of ultralow loss reversible phase-change materials for photonic integrated circuits: Sb2s3 and sb2se3,” Advanced functional materials, vol. 30, no. 36, p. 2002447, 2020.
- [18]M. Delaney, I. Zeimpekis, H. Du, X. Yan, M. Banakar, D. J. Thomson, D. W. Hewak, and O. L. Muskens, “Nonvolatile programmable silicon photonics using an ultralow-loss sb2se3 phase change material,” Science Advances, vol. 7, no. 25, p. eabg3500, 2021.
- [19]R. Chen, Z. Fang, C. Perez, F. Miller, K. Kumari, A. Saxena, J. Zheng, S. J. Geiger, K. E. Goodson, and A. Majumdar, “Non-volatile electrically programmable integrated photonics with a 5-bit operation,” Nature Communications, vol. 14, no. 1, p. 3465, 2023.DOI
- [20]W. Chen, S. Liu, and J. Zhu, “Pixelated non-volatile programmable photonic integrated circuits with 20-level intermediate states,” International Journal of Extreme Manufacturing, vol. 6, no. 3, p. 035501, 2024.
- [21]M. Wei, K. Xu, B. Tang, J. Li, Y. Yun, P. Zhang, Y. Wu, K. Bao, K. Lei, Z. Chen et al., “Monolithic back-end-of-line integration of phase change materials into foundry-manufactured silicon photonics,” Nature Communications, vol. 15, no. 1, p. 2786, 2024.
- [22]D. E. Zelmon, D. L. Small, and D. Hunt, “Infrared corrected sellmeier coefficients for congruently grown lithium niobate and 5 mol.% magnesium oxide–doped lithium niobate,” Journal of the Optical Society of America B, vol. 14, no. 12, pp. 3319–3322, 1997.
- [23]Y. Gutiérrez, A. P. Ovyn, G. Santos, D. Juan, S. A. Rosales, J. Junquera, P. García-Fernández, S. Dicorato, M. M. Giangregorio, E. Dilonardo et al., “Interlaboratory study on sb2s3 interplay between structure, dielectric function, and amorphous-to-crystalline phase change for photonics,” Iscience, vol. 25, no. 6, 2022.
- [24]Flexcompute Inc., “Tidy3D: FDTD and mode solver for electromagnetic simulation,” Computer software, version 2.11.2, 2026, https://www.flexcompute.com/tidy3d/. Source:https://github.com/flexcompute/tidy3d.
- [25]I. M. Sobol, “Distribution of points in a cube and approximate evaluation of integrals,” USSR Computational mathematics and mathematical physics, vol. 7, pp. 86–112, 1967.
- [26]P. Bienstman and R. Baets, “Optical modelling of photonic crystals and vcsels using eigenmode expansion and perfectly matched layers,” Optical and Quantum Electronics, vol. 33, no. 4, pp. 327–341, 2001.
- [27]A. W. Snyder, J. D. Love et al., Optical waveguide theory. Chapman and hall, London, 1983, vol. 175.
- [28]R. P. Feynman, “Forces in molecules,” Physical review, vol. 56, no. 4, p. 340, 1939.
- [29]T. Visser, B. DeMeulenaere, J. Haes, D. Lenstra, R. Baets, and H. Blok, “Confinement and modal gain in dielectric waveguides,” Journal of lightwave technology, vol. 14, no. 5, pp. 885–887, 1996.
- [30]J. T. Robinson, K. Preston, K. Painter, and M. Lipson, “First-principle derivation of gain in high-index-contrast waveguides,” Optics express, vol. 16, no. 21, pp. 16559–16669, 2008.
- [31]N. Sakib, “Theory of extreme optical concentration in all-dielectric waveguides,” 2019.
- [32]S. Johnson, A. Mekis, S. Fan, and J. Joannopoulos, Molding the flow of light, 12 2001, vol. 3.
- [33]M. A. Biot, “General theorems on the equivalence of group velocity and energy transport,” Physical review, vol. 105, no. 4, p. 1129, 1957.
- [34]S. G. Johnson, M. Ibanescu, M. Skorobogaty, O. Weisberg, J. Joannopoulos, and Y. Fink, “Perturbation theory for maxwell’s equations with shifting material boundaries,” Physical review E, vol. 65, no. 6, p. 066611, 2002.
- [35]L. Lu, R. Pestourie, W. Yao, Z. Wang, F. Verdugo, and S. G. Johnson, “Physics-informed neural networks with hard constraints for inverse design,” SIAM Journal on Scientific Computing, vol. 43, no. 6, pp. B1105–B1132, 2021.
- [36]G. E. Karniadakis, I. G. Kevrekidis, L. Lu, P. Perdikaris, S. Wang, and L. Yang, “Physics-informed machine learning,” Nature Reviews Physics, vol. 3, no. 6, pp. 422–440, 2021.
- [37]M. Raissi, P. Perdikaris, and G. E. Karniadakis, “Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations,” Journal of Computational physics, vol. 378, pp. 686–707, 2019.
- [38]Y. Chen, L. Lu, G. E. Karniadakis, and L. Dal Negro, “Physics-informed neural networks for inverse problems in nano-optics and metamaterials,” Optics express, vol. 28, no. 8, pp. 11618–11633, 2020.
- [39]W. M. Czarnecki, S. Osindero, M. Jaderberg, G. Swirszcz, and R. Pascanu, “Solbolev training for neural networks,” in Neural Information Processing Systems, 2017. [Online]. Available: https://api.semanticscholar.org/CorpusID:21596346
- [40]H. S. Ünal and A. C. Durgun, “A physics-aware neural network for effective refractive index prediction of photonic waveguides: Hs ünal, ac durgun,” Optical and Quantum Electronics, vol. 57, no. 1, p. 107, 2025.
- [41]B. Lakshminarayanan, A. Pritzel, and C. Blundell, “Simple and scalable predictive uncertainty estimation using deep ensembles,” in Neural Information Processing Systems, 2016. [Online]. Available: https://api.semanticscholar.org/CorpusID:6294674
- [42]D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” 2014. [Online]. Available: https://api.semanticscholar.org/CorpusID:6628106
- [43]D. C. Liu and J. Nocedal, “On the limited memory bfgs method for large scale optimization,” Mathematical programming, vol. 45, no. 1, pp. 503–528, 1989.
- [44]C. Rasmussen and C. Williams, Gaussian Processes for Machine Learning, 11 2005.
- [45]J. H. Friedman, “Greedy function approximation: a gradient boosting machine,” Annals of statistics, pp. 1189–1232, 2001.
- [46]C. Laprais, C. Zrounba, J. Bouvier, N. Blanchard, M. Bugnet, A. Gassenq, Y. Gutiérrez, S. Vazquez-Miranda, S. Espinoza, P. Thiesen et al., “Reversible single-pulse laser-induced phase change of sb2s3 thin films: Multi-physics modeling and experimental demonstrations,” Advanced optical materials, vol. 12, no. 28, p. 2401214, 2024.
- [47]W. M. Haynes, CRC handbook of chemistry and physics. CRC press, 2016.
- [48]F. Hoff, J. Pries, J. Köttgen, P. Lucas, and M. Wuttig, “Controlling the crystallization kinetics of low loss phase change material sb2s3,” Advanced physics research, vol. 4, no. 9, p. 2500005, 2025.