pith. sign in

arxiv: 2604.02184 · v2 · pith:3EWBKBATnew · submitted 2026-04-02 · 💻 cs.LG

Neural-network methods for two-dimensional finite-source reflector design

Roel Hacking , Lisa Kusch , Koondanibha Mitra , Martijn Anthonissen , Wilbert IJzerman This is my paper

Pith reviewed 2026-05-21 09:25 UTC · model grok-4.3

classification 💻 cs.LG
keywords reflector designneural networksinverse opticsfinite sourceray mappingfar-field distributionoptical optimization2D reflectors
0
0 comments X

The pith

A neural network representing reflector height and optimized with ray-mapping losses designs 2D finite-source reflectors to errors of 2e-5 and 5e-5 in seconds.

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

The paper proposes representing the height of a two-dimensional reflector as the output of a neural network and optimizing its parameters to redirect light from a finite extended source into a prescribed far-field distribution. Two loss functions guide the optimization: a direct change-of-variables loss that uses the closed-form inverse ray map and a mesh-based loss that maps target cells back to the source, which works even for discontinuous sources. Gradients flow through automatic differentiation and a quasi-Newton solver minimizes the objectives. On the main benchmarks the neural method reaches ray-traced normalized mean absolute errors of about 2e-5 and 5e-5 within a few seconds on one GPU, while an adapted deconvolution baseline reaches only 4e-3 and 5e-2 after hundreds of seconds. The formulation also accommodates minimum-height constraints and is discussed for extension to rotationally symmetric and three-dimensional cases.

Core claim

The paper establishes that parameterizing the reflector height with a neural network and minimizing either a direct change-of-variables loss based on the closed-form inverse ray map or a mesh-based loss that remaps target cells to the source, using automatic differentiation and a robust quasi-Newton method, produces designs whose ray-traced normalized mean absolute error on finite-source benchmarks reaches approximately 2e-5 and 5e-5, substantially outperforming an adapted deconvolution pipeline that recovers a monotone map from flux balance, solves an integrating-factor ODE, and applies modified Van Cittert iteration with nonnegativity clipping and ray-traced feedback.

What carries the argument

Neural-network parameterization of reflector height optimized by a direct change-of-variables loss based on the closed-form inverse ray map together with a mesh-based loss that maps target cells back to the source.

If this is right

  • The method supports minimum-height constraints during optimization.
  • The mesh-based loss remains usable for discontinuous source distributions.
  • Accuracy is measured directly by ray-traced normalized mean absolute error across continuous and discontinuous benchmarks.
  • The formulation is discussed for extension to rotationally symmetric and full three-dimensional reflector design via iterative correction schemes.

Where Pith is reading between the lines

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

  • The rapid runtime on consumer GPUs could support repeated design iterations in which the target far-field distribution is refined together with the reflector shape.
  • Automatic differentiation through the ray map suggests the same approach could incorporate additional optical effects such as wavelength dependence or polarization if they are added to the loss.
  • The combination of direct and mesh-based losses may transfer to other inverse mapping problems in optics or acoustics that involve extended sources and prescribed output distributions.

Load-bearing premise

The neural network is expressive enough to approximate the optimal reflector shape and the chosen loss functions are faithful enough to the underlying ray-mapping physics that their minimum produces a reflector whose actual ray-traced performance matches the reported error levels.

What would settle it

Independent ray-tracing of the neural-optimized reflector on either main benchmark that yields a normalized mean absolute error larger than 10^{-4} would show that the claimed accuracy level has not been reached.

Figures

Figures reproduced from arXiv: 2604.02184 by Koondanibha Mitra, Lisa Kusch, Martijn Anthonissen, Roel Hacking, Wilbert IJzerman.

Figure 1
Figure 1. Figure 1: Finite-source-to-far-field reflector system setup. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Finite-source-to-far-field reflector system approximation setup. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Mesh-based procedure for computing a continuous loss for inverse reflector design. (a) The [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Reflectors for the mapping m(s) = s 3 for different values of h computed using Eq. (35). The left plot shows the mapping, the center plot shows the obtained functions u for different values of h, and the right plot shows the corresponding reflectors in the xz-plane. where a(s) = β ′ [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Ground-truth reflector, source distribution, corresponding target distribution, and far-field [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Final reflectors and convergence history for the neural network and deconvolution methods [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Ground-truth reflector, source distribution, corresponding target distribution, and far-field [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Final reflectors and convergence history for the neural network and deconvolution methods [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: NMAE for the neural network and deconvolution methods as a function of minimum [PITH_FULL_IMAGE:figures/full_fig_p014_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: NMAE for the neural network and deconvolution methods as a function of minimum [PITH_FULL_IMAGE:figures/full_fig_p015_10.png] view at source ↗
read the original abstract

We address the inverse problem of designing two-dimensional reflectors that transform light from a finite, extended source into a prescribed far-field distribution. The reflector height is represented by a neural network and optimized with two objective functions: a direct change-of-variables loss based on the closed-form inverse ray map, and a mesh-based loss that maps target cells back to the source and remains usable for discontinuous sources. Gradients are computed by automatic differentiation and minimized with a robust quasi-Newton method. As a baseline, we adapt a deconvolution pipeline built on a simplified finite-source approximation: a one-dimensional monotone map is recovered from flux balance, converted to a reflector by an integrating-factor ODE solve, and embedded in a modified Van Cittert iteration with nonnegativity clipping and ray-traced feedback. Across four benchmarks, covering continuous and discontinuous sources and minimum-height constraints, accuracy is measured by ray-traced normalized mean absolute error. On the two main benchmarks, the neural method reaches errors of about 2e-5 and 5e-5 within a few seconds on one NVIDIA RTX 4090 GPU, compared with 4e-3 and 5e-2 for the deconvolution baseline after several hundred seconds. The results show that the neural formulation is both more accurate and substantially faster, while still supporting practical height constraints. We also discuss extensions to rotationally symmetric and full three-dimensional reflector design through iterative correction schemes.

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.

Referee Report

2 major / 2 minor

Summary. The paper addresses the inverse design of two-dimensional reflectors for finite extended light sources to achieve prescribed far-field distributions. The reflector height is parameterized by a neural network and optimized using two loss functions: a change-of-variables loss based on the closed-form inverse ray map and a mesh-based loss suitable for discontinuous sources. Optimization uses automatic differentiation and quasi-Newton methods. The approach is benchmarked against a deconvolution baseline on four test cases, demonstrating superior accuracy (ray-traced NMAE of approximately 2e-5 and 5e-5) and speed (seconds vs. hundreds of seconds) on an NVIDIA RTX 4090 GPU.

Significance. Should the central claims hold, this represents a notable advance in applying neural methods to optical inverse problems. The reported improvements in both accuracy and runtime, combined with support for practical constraints like minimum height, suggest potential for broader adoption in reflector and lens design. The discussion of extensions to rotationally symmetric and 3D cases adds to its utility. The use of standard tools like AD makes the method accessible.

major comments (2)
  1. [Loss formulation] Loss formulation section: the mesh-based loss maps target cells back to the source via a discrete approximation. For discontinuous sources this discretization error is not shown to vanish at a rate that guarantees the independent ray-traced NMAE reaches the reported 2e-5 / 5e-5 levels; a convergence plot versus mesh resolution (or an explicit bound) is needed to confirm the surrogate is faithful.
  2. [Results] Results, benchmark tables: the deconvolution baseline reports 4e-3 and 5e-2 after several hundred seconds, but the precise number of Van Cittert iterations, the exact finite-source approximation used inside the loop, and whether the same ray-tracer is employed for both methods must be stated explicitly so that the speed/accuracy gap can be reproduced.
minor comments (2)
  1. [Abstract] Abstract: only two of the four benchmarks are quantified; briefly state the NMAE values obtained on the remaining two (continuous/discontinuous, with/without height constraint) for completeness.
  2. [Implementation] Implementation: network architecture (depth, width, activation), quasi-Newton tolerances, and exact form of the minimum-height penalty should be listed or referenced to a supplementary file to support reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading of our manuscript and the constructive comments. We address each major comment below and describe the revisions we will make.

read point-by-point responses

  1. Referee: [Loss formulation] Loss formulation section: the mesh-based loss maps target cells back to the source via a discrete approximation. For discontinuous sources this discretization error is not shown to vanish at a rate that guarantees the independent ray-traced NMAE reaches the reported 2e-5 / 5e-5 levels; a convergence plot versus mesh resolution (or an explicit bound) is needed to confirm the surrogate is faithful.

    Authors: We agree that an explicit demonstration of convergence for the mesh-based loss on discontinuous sources would strengthen the validation. In the revised manuscript we will add a convergence plot of the independent ray-traced NMAE versus mesh resolution for the two discontinuous-source benchmarks. Preliminary checks confirm that the reported NMAE values of approximately 2e-5 and 5e-5 are reached and stabilize once the mesh resolution exceeds the values used in our experiments. revision: yes

  2. Referee: [Results] Results, benchmark tables: the deconvolution baseline reports 4e-3 and 5e-2 after several hundred seconds, but the precise number of Van Cittert iterations, the exact finite-source approximation used inside the loop, and whether the same ray-tracer is employed for both methods must be stated explicitly so that the speed/accuracy gap can be reproduced.

    Authors: We acknowledge that additional implementation details for the baseline are needed for full reproducibility. The revised manuscript will state the exact number of Van Cittert iterations, the precise one-dimensional finite-source approximation employed inside the loop, and confirm that the identical ray-tracer is used to evaluate both the neural-network designs and the baseline. These clarifications will allow direct reproduction of the reported accuracy and runtime figures. revision: yes

Circularity Check

0 steps flagged

No circularity: performance claims rest on independent ray-traced NMAE benchmark

full rationale

The paper optimizes a neural-network reflector height via two explicitly defined surrogate losses (change-of-variables using closed-form inverse map, and mesh-based cell mapping) and then reports accuracy on a separate ray-traced normalized mean absolute error metric. This evaluation is performed after optimization and compared against an independently implemented deconvolution baseline; neither the final metric nor the reported error values are defined in terms of the optimized losses or any self-citation. The derivation chain therefore remains self-contained against external benchmarks with no reduction of outputs to inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract supplies no explicit free parameters, mathematical axioms, or newly postulated physical entities; the method rests on standard neural-network approximation power and geometric optics ray tracing.

pith-pipeline@v0.9.0 · 5798 in / 1234 out tokens · 54120 ms · 2026-05-21T09:25:53.099348+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

35 extracted references · 35 canonical work pages · 2 internal anchors

  1. [1]

    Welford W T and Winston R 1989High Collection Nonimaging Optics(San Diego: Academic Press)

    work page

  2. [2]

    Winston R, Mi˜ nano J C and Ben´ ıtez P 2004Nonimaging Optics(Elsevier Academic Press) ISBN 9780127597515

    work page

  3. [3]

    Chaves J 2015Introduction to Nonimaging Optics2nd ed (CRC Press) ISBN 9781482206739

    work page

  4. [4]

    B¨ osel C and Gross H 2019Applied Optics582713–2724

    work page

  5. [5]

    Wei S, Zhu Z, Li W and Ma D 2021Optics Letters462770–2773

    work page

  6. [6]

    Glimm T and Oliker V I 2003Journal of Mathematical Sciences1174096–4108

    work page

  7. [7]

    Caffarelli L A, Kochengin S and Oliker V 2007ESAIM: Control, Optimisation and Calculus of Variations1393–106

    work page

  8. [8]

    Romijn L B, ten Thije Boonkkamp J H M and IJzerman W L 2019Journal of the Optical Society of America A361926–1939

    work page 1939

  9. [9]

    Romijn L B, ten Thije Boonkkamp J H M and IJzerman W L 2020Journal of Computational Physics408109283

    work page

  10. [10]

    Benamou J D, Chazareix G, IJzerman W and Rukhaia G 2022Journal of Computational Physics456111032 16 IOP PublishingJournalvv(yyyy) aaaaaa Authoret al

    work page

  11. [11]

    B¨ osel C and Gross H 2016Optics Express2414271–14282

    work page

  12. [12]

    B¨ osel C, Worku N G and Gross H 2017Applied Optics563679–3688

    work page

  13. [13]

    Sorgato S, Chaves J, Thienpont H and Duerr F 2019Optica6966–971

    work page

  14. [14]

    SPIEvol 5185 pp 18–29

    Ben´ ıtez P, Mi˜ nano J C, Blen J, Mohedano R, Chaves J, Dross O, Hern´ andez M,´Alvarez J L and Falicoff W 2004 SMS design method in 3D geometry: examples and applicationsProc. SPIEvol 5185 pp 18–29

    work page 2004

  15. [15]

    Wu R, Huang C Y, Zhu X, Cheng H N and Liang R 2016Optica3840–843

    work page

  16. [16]

    Zhu Z, Wei S, Fan Z and Ma D 2022Optics Express3011524–11535

    work page

  17. [17]

    SPIEvol 7423 p 742302

    Fournier F R, Cassarly W J and Rolland J P 2009 Designing freeform reflectors for extended sourcesProc. SPIEvol 7423 p 742302

    work page 2009

  18. [18]

    van Cittert P H 1931Zeitschrift f¨ ur Physik69298–303

    work page

  19. [19]

    Burger H C and van Cittert P H 1932Zeitschrift f¨ ur Physik79722–730

    work page

  20. [20]

    Richardson W H 1972Journal of the Optical Society of America6255–59

    work page

  21. [21]

    Lucy L B 1974The Astronomical Journal79745–754

    work page

  22. [22]

    Hill N R and Ioup G E 1976Journal of the Optical Society of America66487–489

    work page

  23. [23]

    Jansson P A 1997Deconvolution of Images and Spectra2nd ed (Academic Press) ISBN 0123802229

    work page

  24. [24]

    thesis Eindhoven University of Technology Eindhoven, The Netherlands

    Kronberg V C E 2024Inverse freeform reflector design with a scattering surfacePh.D. thesis Eindhoven University of Technology Eindhoven, The Netherlands

    work page

  25. [25]

    Heemels A, de Koning B, Adam A and M¨ oller M 2024Optics Express329730–9746

    work page

  26. [26]

    de Koning B, Heemels A, Adam A and M¨ oller M 2023arXiv preprint arXiv:2302.12031URL https://arxiv.org/abs/2302.12031

    work page arXiv

  27. [27]

    Gannon C and Liang R 2019arXiv preprint arXiv:1903.11166URL https://arxiv.org/abs/1903.11166

    work page internal anchor Pith review Pith/arXiv arXiv 1903

  28. [28]

    Zhou Y, Zhang S, Chen H and Liang R 2024Optics Express326350–6366

    work page

  29. [29]

    Hacking R, Kusch L, Mitra K, Anthonissen M and IJzerman W 2025Journal of Computational Mathematics and Data Science15100119

    work page

  30. [30]

    Urb´ an J F, Stefanou P and Pons J A 2025Journal of Computational Physics523113656

    work page

  31. [31]

    Wang S, Teng Y and Perdikaris P 2021SIAM Journal on Scientific Computing43 A3055–A3081

    work page

  32. [32]

    Moritz P, Nishihara R and Jordan M 2016 A linearly-convergent stochastic L-BFGS algorithm Artificial Intelligence and Statistics(PMLR) pp 249–258

    work page 2016

  33. [33]

    Kingma D P and Ba J 2017 Adam: A method for stochastic optimization (Preprint 1412.6980) URLhttps://arxiv.org/abs/1412.6980

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  34. [34]

    Jnini A, Vella F and Zeinhofer M 2024 Gauss-Newton natural gradient descent for physics-informed computational fluid dynamics (Preprint2402.10680) URL https://arxiv.org/abs/2402.10680

    work page arXiv 2024

  35. [35]

    Kreikemeyer J N and Andelfinger P 2023IEEE Access11143190–143211 17 IOP PublishingJournalvv(yyyy) aaaaaa Authoret al A Ray intersection guarantee We show that the reflector parameterization of Section 2.1 ensures that every emitted ray intersects the reflector curve, regardless of the choice of height functionu, provided only thatuis continuous and strict...

    work page