Machine-learning surrogate model for one-dimensional GaAs/Al$_{0.3}$Ga$_{0.7}$As distributed Bragg reflector spectra

Mehdi Ouslim

arxiv: 2606.08108 · v2 · pith:UR2VJTMFnew · submitted 2026-06-06 · ⚛️ physics.optics · cond-mat.mtrl-sci

Machine-learning surrogate model for one-dimensional GaAs/Al_(0.3)Ga_(0.7)As distributed Bragg reflector spectra

Mehdi Ouslim This is my paper

Pith reviewed 2026-06-27 19:26 UTC · model grok-4.3

classification ⚛️ physics.optics cond-mat.mtrl-sci

keywords Gaussian processsurrogate modeldistributed Bragg reflectorreflectance spectrummachine learningtransfer matrix methodphotonicsuncertainty quantification

0 comments

The pith

A Gaussian process surrogate predicts DBR reflectance spectra seventy times faster than transfer-matrix calculations while supplying calibrated uncertainty.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper trains a Gaussian process model on 1500 transfer-matrix simulations of GaAs/Al0.3Ga0.7As distributed Bragg reflectors to predict their normal-incidence reflectance spectra. Principal component analysis reduces each spectrum to 26 components, after which a separate Gaussian process is fit to every component. The resulting surrogate evaluates a full spectrum in 4.4 milliseconds instead of the 308 milliseconds required by direct transfer-matrix evaluation. On held-out test points the model reports an RMSE of 0.085 and shows that its 95 percent prediction bands contain 98.9 percent of the residuals. The work therefore supplies a fast, uncertainty-aware tool for exploring DBR layer-thickness combinations.

Core claim

A Gaussian-process surrogate model trained on principal components of 1500 Latin-hypercube transfer-matrix simulations predicts the normal-incidence reflectance spectrum of one-dimensional GaAs/Al0.3Ga0.7As distributed Bragg reflectors. On a held-out test set the model attains RMSE = 0.085 and R² = 0.276; its 95 percent prediction bands cover 98.9 percent of the observed residuals. Inference requires 4.4 ms per spectrum, producing a roughly seventy-fold reduction in wall-clock time relative to the original transfer-matrix calculation and thereby enabling rapid design-space exploration.

What carries the argument

The Gaussian process regressor fitted independently to each of the 26 principal components of the reflectance spectrum, which reconstructs the full output while returning per-wavelength uncertainty estimates.

If this is right

Designers can scan large regions of DBR parameter space in minutes rather than hours.
The supplied uncertainty bands identify regions of the design space where predictions are most trustworthy.
The surrogate can be inserted directly into iterative optimization loops that adjust layer thicknesses to meet target reflectance profiles.
The same principal-component-plus-Gaussian-process pipeline can be retrained for other multilayer optical stacks once new transfer-matrix data are available.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The approach could be combined with the Random Forest baseline already tested in the paper to trade some uncertainty quantification for higher point-wise accuracy.
Real-time graphical interfaces that let a user drag layer thicknesses and instantly see the updated spectrum become feasible once the model is embedded in design software.
Retraining the surrogate on spectra that include oblique incidence or different material systems would extend its utility to a wider class of photonic components.

Load-bearing premise

The 1500 Latin-hypercube samples adequately cover the ranges of layer thicknesses and wavelengths that matter for practical DBR designs.

What would settle it

A fresh set of transfer-matrix simulations performed at parameter values outside the original Latin-hypercube bounds, where the Gaussian process predictions fall outside the reported 95 percent bands on a substantial fraction of wavelengths.

Figures

Figures reproduced from arXiv: 2606.08108 by Mehdi Ouslim.

**Figure 1.** Figure 1: Parity plots of GP-predicted versus TMM reflectance at 950, 1000, and 1050 nm. Points are coloured by Nperiods; the dashed line indicates perfect agreement. Per-wavelength RMSE values are annotated. Funding No external funding. Disclosures The authors declare no conflicts of interest. Data availability Dataset and code are available at https://github. com/mehdiouslim-hash/ai-laue-research. References [1] … view at source ↗

**Figure 3.** Figure 3: Full-spectrum comparison for 12 randomly selected test samples. Black lines: TMM ground truth. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

**Figure 4.** Figure 4: Scatter plots of GP-predicted versus TMM stopband centre wavelength (left) and peak reflectance (right). RMSE and R2 are annotated on each panel. 0 100 200 300 400 500 Training set size N 0.02 0.04 0.06 0.08 0.10 T e st R M S E in R GP (PCA) Random Forest Target RMSE = 0.02 [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

**Figure 5.** Figure 5: Test RMSE as a function of training set size for GP (PCA) and Random Forest models. The horizontal dotted line marks the 0.02 target RMSE. Neither model has saturated at N = 500, motivating larger datasets. [19] F. Davoodi, “From bound states to quantum spin models: chiral coherent dynamics in topological photonic rings,” Nanophotonics 14, 4397–4409 (2025). 7 [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗

read the original abstract

We present a Gaussian-process (GP) surrogate model for the normal-incidence reflectance spectrum of one-dimensional GaAs/Al$_{0.3}$Ga$_{0.7}$ distributed Bragg reflectors (DBRs). A Latin-hypercube dataset of 1500 transfer-matrix-method (TMM) simulations is used to train and evaluate the model. Principal component analysis reduces the spectral output to 26 components; one GP is fitted per component. On a held-out test set the GP achieves RMSE=0.085 and $R^2=0.276$, while a Random Forest baseline reaches RMSE=0.065 and $R^2=0.572$. GP inference is 4.4 ms per spectrum compared with ~308 ms for TMM, yielding a ~70x speedup. Uncertainty calibration shows that the GP 95% prediction band covers 98.9% of test residuals. These results establish a rapid surrogate for DBR design-space exploration.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The GP-PCA surrogate delivers a 70x speedup on DBR spectra but only reaches R²=0.276, well below the random forest baseline, so the practical value looks limited.

read the letter

The paper fits independent Gaussian processes to the first 26 principal components of reflectance spectra from 1500 Latin-hypercube TMM runs on GaAs/Al0.3Ga0.7As DBRs. On held-out data it reports RMSE=0.085, R²=0.276, and 95% prediction intervals that cover 98.9% of residuals, with inference at 4.4 ms versus 308 ms for TMM.

The work does a few things cleanly. It uses a standard but reproducible sampling scheme, keeps test data separate, supplies a direct timing comparison, and checks calibration of the uncertainty estimates. Those steps are straightforward and worth noting.

The numbers still raise a clear problem. An R² of 0.276 means the model accounts for only about a quarter of the spectral variance after PCA reconstruction. The random forest baseline reaches 0.572 on the same split, so the GP choice does not appear to be the stronger option here. If the goal is usable design-space exploration, point predictions that miss most of the variation will not substitute reliably for TMM. The abstract presents the result as establishing a rapid surrogate, yet the fit quality undercuts that claim.

No information appears on how the 26 components were selected, whether any hyperparameter search was done, or what preprocessing was applied to the spectra. Those omissions make it hard to tell whether the reported performance is the best achievable or simply the first attempt.

This is a narrow methods note aimed at people already running DBR optimizations who might want a quick drop-in approximation. Readers looking for either a methodological advance or a convincingly accurate surrogate will not find enough here. I would not send it for peer review; the gap between the speedup claim and the actual explanatory power is too large to ignore without further work.

Referee Report

3 major / 1 minor

Summary. The manuscript develops a Gaussian-process surrogate for normal-incidence reflectance spectra of one-dimensional GaAs/Al0.3Ga0.7As DBRs. A 1500-point Latin-hypercube sample of TMM simulations is reduced via PCA to 26 components; independent GPs are trained on each component. On a held-out test set the GP reports RMSE = 0.085 and R² = 0.276 (Random Forest baseline: RMSE = 0.065, R² = 0.572), inference time 4.4 ms versus ~308 ms for TMM (~70× speedup), and 95 % prediction bands that cover 98.9 % of residuals. The authors conclude that the surrogate enables rapid DBR design-space exploration.

Significance. A well-calibrated, fast surrogate for DBR spectra would be useful for design optimization. The reported uncertainty calibration is a positive feature, but the low GP R² relative to the Random Forest baseline limits the practical utility of the presented model for design exploration.

major comments (3)

[Abstract] Abstract: the statement that the results 'establish a rapid surrogate for DBR design-space exploration' is not supported by the reported metrics; the GP R² = 0.276 is substantially lower than the Random Forest R² = 0.572, indicating that the surrogate explains only ~28 % of the spectral variance and therefore cannot reliably support design-space exploration.
[Methods] Methods / Results: no details are given on GP hyperparameter selection, kernel choice, or the procedure used to determine the number of retained PCA components (26). These choices directly affect the reported R² and RMSE and must be documented to allow assessment of robustness.
[Results] Results: the claim of adequate sampling by the 1500-point Latin-hypercube design is not accompanied by any convergence test (e.g., learning curves or additional validation points outside the original hypercube), leaving open whether the low GP R² arises from insufficient coverage of the design space.

minor comments (1)

[Abstract] The abstract and main text should explicitly state the wavelength range and number of spectral points used in the TMM simulations and PCA.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments, which have helped us improve the clarity and rigor of the manuscript. We address each major comment below and have made revisions where the concerns are valid.

read point-by-point responses

Referee: [Abstract] Abstract: the statement that the results 'establish a rapid surrogate for DBR design-space exploration' is not supported by the reported metrics; the GP R² = 0.276 is substantially lower than the Random Forest R² = 0.572, indicating that the surrogate explains only ~28 % of the spectral variance and therefore cannot reliably support design-space exploration.

Authors: We agree that the original abstract phrasing overstated the surrogate's utility for design-space exploration given the modest R². While the GP provides calibrated uncertainty estimates (a feature absent from the Random Forest baseline) and a ~70× speedup, the accuracy is indeed limited. We have revised the abstract to describe the model as providing a rapid approximation with uncertainty quantification, while explicitly noting the R² relative to the baseline and its implications for practical use. revision: yes
Referee: [Methods] Methods / Results: no details are given on GP hyperparameter selection, kernel choice, or the procedure used to determine the number of retained PCA components (26). These choices directly affect the reported R² and RMSE and must be documented to allow assessment of robustness.

Authors: We have added a dedicated paragraph in the Methods section specifying the kernel (squared-exponential with automatic relevance determination), hyperparameter optimization via marginal-likelihood maximization, and the PCA retention criterion (components explaining cumulative variance ≥99%, yielding 26 components). These details allow reproducibility and assessment of robustness. revision: yes
Referee: [Results] Results: the claim of adequate sampling by the 1500-point Latin-hypercube design is not accompanied by any convergence test (e.g., learning curves or additional validation points outside the original hypercube), leaving open whether the low GP R² arises from insufficient coverage of the design space.

Authors: We have added learning curves in the revised Results section, obtained by training on random subsamples of the 1500-point set (sizes 100–1500). Performance plateaus beyond ~1200 points, supporting that the design provides adequate coverage within the sampled hypercube. External validation points outside the hypercube would require new TMM runs, which we have not performed. revision: partial

Circularity Check

0 steps flagged

No circularity: performance metrics measured on independent held-out TMM data

full rationale

The paper generates a Latin-hypercube dataset of 1500 TMM simulations, applies PCA to reduce spectra to 26 components, fits one GP per component, and reports RMSE, R², and uncertainty calibration exclusively on a held-out test set of TMM simulations never seen during training or hyperparameter selection. No equation or claim reduces a prediction to a fitted input by construction, no self-citation is invoked as a uniqueness theorem or load-bearing premise, and the speedup comparison is a direct wall-clock measurement against the same TMM code. The derivation chain is therefore a standard empirical ML workflow whose central claims remain falsifiable against external TMM runs.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The approach rests on standard assumptions of Gaussian process regression and the representativeness of the Latin-hypercube sample; no new physical entities or ad-hoc axioms are introduced beyond routine ML practice.

free parameters (1)

Number of PCA components = 26
Spectral output reduced to 26 components; selection criterion not stated in abstract.

axioms (1)

domain assumption Gaussian process assumptions (stationarity, appropriate kernel) hold for the PCA-reduced spectral components
Invoked by fitting one independent GP per component.

pith-pipeline@v0.9.1-grok · 5708 in / 1050 out tokens · 23889 ms · 2026-06-27T19:26:46.024434+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

19 extracted references

[1]

B. E. A. Saleh and M. C. Teich,Fundamentals of Photonics, 3rd ed. (Wiley, 2019)

2019
[2]

In- terfacing single photons and single quantum dots with photonic nanostructures,

P. Lodahl, S. Mahmoodian, and S. Stobbe, “In- terfacing single photons and single quantum dots with photonic nanostructures,” Rev. Mod. Phys. 87, 347 (2015)

2015
[3]

Born and E

M. Born and E. Wolf,Principles of Optics, 7th ed. (Cambridge University Press, 1999)

1999
[4]

E. D. Palik, ed.,Handbook of Optical Constants of Solids(Academic Press, 1985). 50 75 100 125 150 175 200 tGaAs (nm) 60 80 100 120 140 160 180 200tAlGaAs (nm) 0.05 0.10 0.15 0.20 MAE in R Figure 2: Mean absolute error per test sample plot- ted over the (t GaAs, tAlGaAs) plane. Error is broadly distributed with no systematic failure region

1985
[5]

The refractive index of Al xGa1−xAs below the band gap,

S. Gehrsitz, F. K. Reinhart, C. Gourgon, N. Her- res, A. Vonlanthen, and H. Sigg, “The refractive index of Al xGa1−xAs below the band gap,” J. Appl. Phys.87, 7825 (2000)

2000
[6]

Multilayer optical calculations,

S. J. Byrnes, “Multilayer optical calculations,” arXiv:1603.02720 (2016);tmmPython package v0.1.8 (2020)

arXiv 2016
[7]

A comparison of three methods for selecting values of input variables in the analysis of output from a computer code,

M. D. McKay, R. J. Beckman, and W. J. Conover, “A comparison of three methods for selecting values of input variables in the analysis of output from a computer code,” Technometrics 21, 239 (1979)

1979
[8]

Machine learning inverse problem for topological photonics,

L. Pilozzi, F. A. Farrelly, G. Marcucci, and C. Conti, “Machine learning inverse problem for topological photonics,” Commun. Phys.1, 57 (2018)

2018
[9]

Generative model for the inverse design of metasurfaces,

Z. Liu, D. Zhu, S. P. Rodrigues, K.-T. Lee, and W. Cai, “Generative model for the inverse design of metasurfaces,” Nano Lett.18, 6570 (2018)

2018
[10]

Plasmonic nanos- tructure design and characterization via Deep Learning,

I. Malkiel, M. Mrejen, A. Nagler, U. Arieli, L. Wolf, and H. Suchowski, “Plasmonic nanos- tructure design and characterization via Deep Learning,” Light Sci. Appl.7, 60 (2018). 5 900 950 1000 1050 1100 0.00 0.25 0.50 0.75 1.00R N=9 tG=159 tA=91 GP ±1 TMM GP mean 900 950 1000 1050 1100 0.00 0.25 0.50 0.75 1.00 N=12 tG=182 tA=51 900 950 1000 1050 1100 0.00 ...

2018
[11]

C. E. Rasmussen and C. K. I. Williams,Gaus- sian Processes for Machine Learning(MIT Press, 2006)

2006
[12]

A unifying view of sparse approximate Gaussian process regression,

J. Qui˜ nonero-Candela and C. E. Rasmussen, “A unifying view of sparse approximate Gaussian process regression,” J. Mach. Learn. Res.6, 1939 (2005)

1939
[13]

Gaussian processes for big data,

J. Hensman, N. Fusi, and N. D. Lawrence, “Gaussian processes for big data,” inProc. UAI (2013), pp. 282–290

2013
[14]

Deep kernel learning,

A. G. Wilson, Z. Hu, R. Salakhutdinov, and E. P. Xing, “Deep kernel learning,” inProc. AIS- TATS(2016), pp. 370–378

2016
[15]

Confor- mal prediction: A gentle introduction,

A. N. Angelopoulos and S. Bates, “Confor- mal prediction: A gentle introduction,” Found. Trends Mach. Learn.16, 494 (2023)

2023
[16]

Finite- difference time-domain modeling of monolayer graphene devices at near-infrared wavelengths,

F. Davoodi and N. Granpayeh, “Finite- difference time-domain modeling of monolayer graphene devices at near-infrared wavelengths,” J. Nanophotonics11, 046008 (2017)

2017
[17]

Near-infrared absorbers based on the heterostructures of two- dimensional materials,

F. Davoodi and N. Granpayeh, “Near-infrared absorbers based on the heterostructures of two- dimensional materials,” Appl. Opt.57, 1358– 1366 (2018)

2018
[18]

Active physics-informed deep learning: surrogate modeling for nonplanar wavefront excitation of topological nanophotonic devices,

F. Davoodi, “Active physics-informed deep learning: surrogate modeling for nonplanar wavefront excitation of topological nanophotonic devices,” Nano Lett.25, 768–775 (2025). 6 900 950 1000 1050 1100 TMM centre (nm) 900 925 950 975 1000 1025 1050 1075 1100GP centre (nm) RMSE=69.389 R2=0.1043 0.3 0.4 0.5 0.6 0.7 0.8 0.9 TMM Rpeak 0.3 0.4 0.5 0.6 0.7 0.8 0.9...

2025
[19]

From bound states to quan- tum spin models: chiral coherent dynamics in topological photonic rings,

F. Davoodi, “From bound states to quan- tum spin models: chiral coherent dynamics in topological photonic rings,” Nanophotonics14, 4397–4409 (2025). 7

2025

[1] [1]

B. E. A. Saleh and M. C. Teich,Fundamentals of Photonics, 3rd ed. (Wiley, 2019)

2019

[2] [2]

In- terfacing single photons and single quantum dots with photonic nanostructures,

P. Lodahl, S. Mahmoodian, and S. Stobbe, “In- terfacing single photons and single quantum dots with photonic nanostructures,” Rev. Mod. Phys. 87, 347 (2015)

2015

[3] [3]

Born and E

M. Born and E. Wolf,Principles of Optics, 7th ed. (Cambridge University Press, 1999)

1999

[4] [4]

E. D. Palik, ed.,Handbook of Optical Constants of Solids(Academic Press, 1985). 50 75 100 125 150 175 200 tGaAs (nm) 60 80 100 120 140 160 180 200tAlGaAs (nm) 0.05 0.10 0.15 0.20 MAE in R Figure 2: Mean absolute error per test sample plot- ted over the (t GaAs, tAlGaAs) plane. Error is broadly distributed with no systematic failure region

1985

[5] [5]

The refractive index of Al xGa1−xAs below the band gap,

S. Gehrsitz, F. K. Reinhart, C. Gourgon, N. Her- res, A. Vonlanthen, and H. Sigg, “The refractive index of Al xGa1−xAs below the band gap,” J. Appl. Phys.87, 7825 (2000)

2000

[6] [6]

Multilayer optical calculations,

S. J. Byrnes, “Multilayer optical calculations,” arXiv:1603.02720 (2016);tmmPython package v0.1.8 (2020)

arXiv 2016

[7] [7]

A comparison of three methods for selecting values of input variables in the analysis of output from a computer code,

M. D. McKay, R. J. Beckman, and W. J. Conover, “A comparison of three methods for selecting values of input variables in the analysis of output from a computer code,” Technometrics 21, 239 (1979)

1979

[8] [8]

Machine learning inverse problem for topological photonics,

L. Pilozzi, F. A. Farrelly, G. Marcucci, and C. Conti, “Machine learning inverse problem for topological photonics,” Commun. Phys.1, 57 (2018)

2018

[9] [9]

Generative model for the inverse design of metasurfaces,

Z. Liu, D. Zhu, S. P. Rodrigues, K.-T. Lee, and W. Cai, “Generative model for the inverse design of metasurfaces,” Nano Lett.18, 6570 (2018)

2018

[10] [10]

Plasmonic nanos- tructure design and characterization via Deep Learning,

I. Malkiel, M. Mrejen, A. Nagler, U. Arieli, L. Wolf, and H. Suchowski, “Plasmonic nanos- tructure design and characterization via Deep Learning,” Light Sci. Appl.7, 60 (2018). 5 900 950 1000 1050 1100 0.00 0.25 0.50 0.75 1.00R N=9 tG=159 tA=91 GP ±1 TMM GP mean 900 950 1000 1050 1100 0.00 0.25 0.50 0.75 1.00 N=12 tG=182 tA=51 900 950 1000 1050 1100 0.00 ...

2018

[11] [11]

C. E. Rasmussen and C. K. I. Williams,Gaus- sian Processes for Machine Learning(MIT Press, 2006)

2006

[12] [12]

A unifying view of sparse approximate Gaussian process regression,

J. Qui˜ nonero-Candela and C. E. Rasmussen, “A unifying view of sparse approximate Gaussian process regression,” J. Mach. Learn. Res.6, 1939 (2005)

1939

[13] [13]

Gaussian processes for big data,

J. Hensman, N. Fusi, and N. D. Lawrence, “Gaussian processes for big data,” inProc. UAI (2013), pp. 282–290

2013

[14] [14]

Deep kernel learning,

A. G. Wilson, Z. Hu, R. Salakhutdinov, and E. P. Xing, “Deep kernel learning,” inProc. AIS- TATS(2016), pp. 370–378

2016

[15] [15]

Confor- mal prediction: A gentle introduction,

A. N. Angelopoulos and S. Bates, “Confor- mal prediction: A gentle introduction,” Found. Trends Mach. Learn.16, 494 (2023)

2023

[16] [16]

Finite- difference time-domain modeling of monolayer graphene devices at near-infrared wavelengths,

F. Davoodi and N. Granpayeh, “Finite- difference time-domain modeling of monolayer graphene devices at near-infrared wavelengths,” J. Nanophotonics11, 046008 (2017)

2017

[17] [17]

Near-infrared absorbers based on the heterostructures of two- dimensional materials,

F. Davoodi and N. Granpayeh, “Near-infrared absorbers based on the heterostructures of two- dimensional materials,” Appl. Opt.57, 1358– 1366 (2018)

2018

[18] [18]

Active physics-informed deep learning: surrogate modeling for nonplanar wavefront excitation of topological nanophotonic devices,

F. Davoodi, “Active physics-informed deep learning: surrogate modeling for nonplanar wavefront excitation of topological nanophotonic devices,” Nano Lett.25, 768–775 (2025). 6 900 950 1000 1050 1100 TMM centre (nm) 900 925 950 975 1000 1025 1050 1075 1100GP centre (nm) RMSE=69.389 R2=0.1043 0.3 0.4 0.5 0.6 0.7 0.8 0.9 TMM Rpeak 0.3 0.4 0.5 0.6 0.7 0.8 0.9...

2025

[19] [19]

From bound states to quan- tum spin models: chiral coherent dynamics in topological photonic rings,

F. Davoodi, “From bound states to quan- tum spin models: chiral coherent dynamics in topological photonic rings,” Nanophotonics14, 4397–4409 (2025). 7

2025