{"id":"d7603800-51fb-49a0-a354-7201a551f673","arxiv_id":"2502.04719","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A deep-optics pipeline that injects manufacturing and assembly tolerances into differentiable ray tracing and jointly optimizes lens and decoder, improving simulated deblurring robustness by about 2 dB PSNR.","lead":"To make deep optics survive real manufacturing errors, this paper modifies the training ray tracer to simulate lens misalignment, tilt, thickness, and curvature errors during optimization, co-training the lens and the reconstruction network. Under simulated tolerances the co-designed system gains about 2 dB PSNR, but the physical experiment only retrains the decoder on a fixed commercial lens rather than fabricating the optimized design.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Re robustness: gains are measured under the same Gaussian tolerance model used for training; the real-world experiment never fabricates the optimized lens, so design-to-manufacturing transfer is unverified.","rationale":"I read the paper in good faith. The contribution is genuinely useful: it makes tolerances differentiable and integrates them into deep optics training, with reasonable ablations, code release, and Zemax spot-diagram agreement. The reader's conditional verdict is appropriate because the main unresolved risk is empirical, not logical. The weakest link is the tolerance model itself: Sec. 3.1 and Supp. A define a clamped Gaussian, independent per lens, with ranges chosen in Tab. A1; Sec. 4.2 then tests under identical draws, so the 2 dB gain is a self-consistent simulation result. Without a fabricated optimized lens or at least measured tolerance data from the target process, the central claim of closing the design-to-manufacturing gap is not yet supported. The real-world section cannot substitute for this because it uses a stock lens and only retrains the decoder. A secondary numerical check also weakens the abstract's 'over 2dB' phrasing: averaging the two lens improvements in Tab. 2 gives roughly 1.35 dB, not over 2 dB, unless only Lens2 is meant. These concerns are addressable: an out-of-distribution robustness evaluation or a small fabrication run with interferometric and mechanical measurements would settle them. No fundamental mathematical error is apparent, so I do not move the verdict; I keep it CONDITIONAL, reported here as UNCHANGED.","tokens_in":13973,"tokens_out":6392,"duration_ms":71127,"concrete_test":"Fabricate 10-20 copies of the tolerance-optimized and baseline Lens2 with the intended production process; measure each copy's surface radii, central thicknesses, tilts, and decentrations; render PSFs with the paper's ray tracer from these measured parameters and compute DIV2K deblurring PSNR. If the measured deviations differ materially from N(0, max²/9) and the PSNR margin over the baseline drops below or reverses the Tab. 2 gain, the central design-to-manufacturing claim is an artifact of the assumed Gaussian sampler.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that tolerance-aware deep optics preserves deblurring quality under manufacturing/assembly deviations rests on Sec. 3.1 and Supp. A treating every tolerance as an independent clamped Gaussian, θΔ ~ N(0, max²/9), with hand-set ranges in Tab. A1. Sec. 4.2 evaluates by sampling the same distribution 100 times, and Supp. B.2 confirms the evaluation uses the same sampler; hence Tab. 2 measures fit to an assumed model, not to actual fabricated statistics. The real-world experiment (Sec. 4.3) retrains only the decoder on an off-the-shelf lens and never fabricates Lens1 or Lens2, so it does not test the co-designed optics. Additionally, the abstract's 'over 2dB improvement' is not what Tab. 2 shows as printed: Lens1 gains 0.36 dB and Lens2 gains 2.33 dB, averaging about 1.35 dB. This does not make the method wrong, but it makes the headline robustness claim conditional on an unvalidated tolerance model.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a tolerance-aware deep optics framework that incorporates manufacturing and assembly tolerances (decentration, tilt, central thickness, curvature) into differentiable ray tracing, and jointly optimizes the lens design and the computational decoder. The method is evaluated in simulation on two lens designs against a non-tolerance-aware baseline and a Zemax-based tolerance optimization, and with a real-world experiment using an off-the-shelf lens. The main claims are improved deblurring robustness under tolerances, accurate tolerance modeling validated against Zemax, and a design flow that reduces the design-to-manufacturing gap.","tokens_in":14188,"tokens_out":4735,"duration_ms":43086,"significance":"The problem addressed is important: conventional deep optics assumes perfect fabrication, which creates a gap between simulation and physical systems. Incorporating multiple tolerance types into differentiable ray tracing is a sensible and practically relevant extension, and the Zemax comparison (Sec. 4.1) gives evidence that the perturbed ray tracing is physically accurate. The ablations (Secs. 5.1–5.3) are also useful, particularly the finding that joint optics-decoder optimization is needed for robustness. If the tolerance model were validated against real fabricated lenses, the framework would be a solid contribution to computational imaging. The paper also provides a useful analysis of loss terms and sampling numbers.","major_comments":[{"comment":"The abstract and Sec. 4.2 claim 'over 2dB improvement' in average deblurring performance, but the numbers in Tab. 2 show improvements of 0.36 dB for Lens1 (29.61 vs. 29.25) and 2.33 dB for Lens2 (28.08 vs. 25.75), averaging about 1.35 dB. This claim should be corrected to match the reported results, or the statistical basis for a 2 dB figure should be stated explicitly.","section":"Abstract; Sec. 4.2; Tab. 2"},{"comment":"The 'real-world experiment' does not fabricate or test the tolerance-optimized lens designs; it only retrains a decoder on an off-the-shelf lens and applies artificial perturbations to the acquired images. As a result, this experiment does not validate the central claim of design-to-manufacturing transfer. The section should be reframed as a decoder-robustness study, or supplemented with a fabricated-lens experiment that actually measures performance under real manufacturing deviations.","section":"Sec. 4.3"},{"comment":"The tolerance perturbations are defined in Sec. 3.1 and Supp. A as independent clamped Gaussians with hand-set ranges (Tab. A1), and the evaluation in Sec. 4.2 samples from exactly the same distribution used in training. Consequently, the robustness gains in Tab. 2 are in-sample with respect to the assumed error model and do not yet establish robustness under measured manufacturing statistics. The authors should either validate the tolerance model against real fabrication measurements (e.g., measured decentration/tilt distributions of produced lenses) or explicitly qualify the central robustness claim as conditional on the assumed Gaussian model.","section":"Sec. 3.1; Supp. A; Sec. 4.2"}],"minor_comments":[{"comment":"In Eq. (5), the same symbol Pλ,f is used for both the traced ray position and the averaged centroid; please use a distinct notation such as P̄λ,f for the centroid to avoid ambiguity.","section":"Eq. (5)"},{"comment":"Eq. (7) is missing an operator between PSFIdeal and PSFP erb; the intended convolution or correlation should be explicitly written.","section":"Eq. (7)"},{"comment":"The claim that spot-size errors relative to Zemax are '< 1µm' is not backed by a quantitative table; please report the per-field RMS spot error values so the accuracy claim is verifiable.","section":"Sec. 4.1"},{"comment":"Please clarify whether the off-the-shelf lens used in the real-world experiment corresponds to Lens1 or Lens2, and if not, state explicitly that the physical experiment does not use the designed lens parameters.","section":"Sec. 4.3"},{"comment":"The phrase 'the first end-to-end tolerance-aware optimization framework' should be qualified, since Li et al. [22] and Zheng et al. [46] already address fabrication tolerances in deep optics; if the claim is restricted to refractive lenses, this restriction should be stated explicitly.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The core idea is promising and the simulation results are internally consistent, but the gap between the simulated tolerance model and physical fabrication is the main weakness. The overstatement of the 2 dB gain and the mismatch between the real-world experiment and the actual claims need to be fixed before publication. I would ask for either a fabricated-lens experiment or a substantial softening of the real-world validation claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The real contribution here is the first end-to-end treatment of manufacturing tolerances inside differentiable ray tracing for refractive-lens deep optics, co-optimizing the lens and the decoder against sampled decentration, tilt, thickness, and curvature errors. That is a legitimate step forward, and the paper does several things carefully: the rigid-body transforms for decentration/tilt/thickness are clean, the Zemax comparison shows sub-micron agreement in RMS spot size under perturbations, and the two auxiliary losses (Spot and PSF similarity) are well motivated—the ablation in Table 4 shows training without them collapses. The simulated deblurring gains on Lens2 are real, and the comparison against Zemax-only tolerance optimization is a nice illustration of the optics/decoder mismatch problem.\n\nNow the soft spots. The robustness claim is in-sample by construction: the tolerances are drawn from the same Gaussian model (Sec. 3.1, Tab. A1) used to train, and evaluation samples that same distribution. So Table 2 demonstrates fit to an assumed error model, not transfer to actual manufacturing statistics. The real-world experiment does not test the co-designed lens at all; it retrains only a decoder on an off-the-shelf lens with artificial perturbations. That validates decoder adaptability, not the tolerance-optimized optics. And the abstract's \"over 2dB improvement\" does not match Table 2 as printed: Lens1 gains 0.36 dB, Lens2 gains 2.33 dB, roughly 1.35 dB on average. Minor but it should be corrected. The yield analysis inherits the same in-sample caveat.\n\nNone of this is fatal to the method. The missing piece is a fabrication study with measured tolerance statistics, which would let the authors compare against actual manufacturing spreads. Absent that, the central claim should be framed as robustness under the assumed tolerance model, with the real-world experiment repositioned as a decoder-only study.\n\nThis deserves a serious referee. The topic matters, the method is novel, and the simulations are consistent. Send it to peer review, but expect the authors to either run a fabrication experiment or narrow the claim.","headline":"First to co-optimize refractive-lens deep optics against manufacturing tolerances in the loop, but the headline robustness claim is in-sample and the '>2dB' gain is overstated.","tokens_in":14722,"tokens_out":2494,"would_cite":false,"duration_ms":22820,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Deep optics usually assumes a perfect lens; this paper claims the first end-to-end framework that instead trains the lens and its reconstruction network against sampled manufacturing and assembly tolerances, recovering more than 2 dB of…","keywords":["deep optics","tolerance-aware optimization","differentiable ray tracing","lens design","manufacturing tolerances","computational imaging","point spread function","image deblurring"],"falsifier":"Fabricate, say, 100 copies of the tolerance-optimized Lens 2, measure their actual decentration, tilt, thickness, and curvature errors, and compare the distribution of measured deblurring PSNR against the simulated 100-sample tolerance test in Table 2; if the measured performance spread matches the conventional design's spread rather than the tolerance-aware design's, the Gaussian independence model and the >2 dB claim are falsified. A cheaper preliminary check is to measure a real factory's tolerance statistics and test whether they fit the assumed normal distribution with the stated ranges.","tokens_in":13772,"feed_emoji":"📷","tokens_out":8210,"duration_ms":76065,"temperature":0.7,"pith_summary":"Deep optics designs a lens and a reconstruction network as one trainable system, but it usually assumes the manufactured lens matches the design exactly. This paper argues that the unavoidable errors of mass production—decentering, tilt, thickness, and curvature deviations on each lens—are precisely what makes fabricated deep-optics cameras underperform, and that these errors should be built into the design process itself. It proposes what it calls the first end-to-end tolerance-aware optimization framework: every training iteration randomly perturbs each lens surface by tolerances sampled from a Gaussian, traces rays through the perturbed system, and back-propagates the reconstruction loss into both the optical parameters and the decoder. The authors claim that the resulting design keeps its deblurring quality under random deviations, with over 2 dB PSNR improvement over conventional deep optics, and that the tolerance-aware design also raises manufacturing yield. The reason this matters is that it targets the design-to-manufacturing gap that keeps deep-optics lenses from being practical in volume production.","feed_headline":"Tolerance-aware training recovers 2 dB in deep optics deblurring","feed_subtitle":"Sampling factory deviations inside ray tracing keeps the lens and decoder sharp after manufacturing.","key_machinery":"The load-bearing object is a tolerance-aware differentiable ray tracer. In each forward pass, the four tolerance types are converted into per-surface spatial transformations—translation for decentration and central thickness, rotation for tilt, plus a curvature offset—and applied as an equivalent coordinate transform on the ray before intersection and refraction, keeping the whole pipeline differentiable. Training then minimizes three terms: a standard image-quality loss, a Spot loss that keeps the traced spot size within a reasonable range, and a PSF similarity loss that penalizes changes in the point spread function under random tolerances. The PSF-similarity term is what couples the optical design to the decoder's expectations, stabilizing the Monte Carlo training and forcing the lens to keep its encoding stable under perturbations.","core_discovery":"The central claim is that explicitly modeling manufacturing and assembly tolerances inside the differentiable ray tracer, and optimizing against them jointly with the decoder, produces a deep-optics system whose performance no longer collapses when the real lens deviates from the nominal design. Concretely, the paper shows that sampling decentration, tilt, central-thickness, and curvature errors per lens and per training iteration, rendering the perturbed point spread function map, and training the optics and reconstruction network together yields more than 2 dB higher average PSNR than the same pipeline trained without tolerances, when both are tested under random tolerances. It also shows that the perturbed ray tracing matches Zemax spot diagrams to within about 1 micrometer in root-mean-square spot size, that a tolerance-aware design has substantially better manufacturing yield, and that optimizing only the optics or only the decoder is not enough: the gain appears only when both are updated together. On the basis of its real-world experiment, the paper further claims that the optimized decoder is more robust to actual perturbations during image acquisition.","pith_inferences":["A direct next test is physical fabrication: the paper's real-world experiment perturbs an off-the-shelf lens and optimizes only the decoder, so it never manufactures the tolerance-optimized lens designs; measuring actual point spread functions and tolerance statistics of fabricated copies of Lens 1 or Lens 2 would confirm or refute whether the Gaussian model and the claimed >2 dB gain survive real","Because the framework treats tolerances as differentiable random perturbations, the same recipe could be transferred to other fabrication errors, such as surface irregularity or refractive-index inhomogeneity, or to non-refractive optics like diffractive elements and metasurfaces where manufacturing errors have different distributions.","The PSF-similarity loss acts as a stability regularizer, so a plausible untested consequence is that it may also improve robustness to defocus, thermal drift, or other perturbations that alter the point spread function, not only assembly tolerances.","If the tolerance model is accurate, the proposed flow suggests an industrial recipe: start from a standard design that already meets image-quality goals, then harden it and its decoder against measured factory tolerance statistics rather than hand-set ranges."],"forward_implications":["Under the paper's tolerance model, a deep-optics lens optimized with this framework keeps an average PSNR above 28 dB when a conventional design drops to around 26 dB on the paper's deblurring test set with random tolerances.","The same designs show higher manufacturing yield: for Lens 1, the tolerance-aware design reaches 26.26 dB PSNR with 90% confidence, versus 24.30 dB for the non-tolerant design.","Tolerance optimization by optics-only metrics, as done with Zemax, can degrade end-to-end deblurring because it ignores the encoder-decoder pairing; the paper's joint optimization avoids this mismatch, as seen for Lens 2 in its Table 2.","Optimizing only the decoder improves its robustness but leaves the design fragile, and optimizing only the optics also fails, so the framework's benefit requires joint optimization of both.","The two-stage recipe—pretrain without tolerances, then tolerance-aware fine-tuning—offers a stable path that protects the pre-trained design's image-quality goals while hardening the system against deviations."],"supporting_citations":[{"why":"Supplies the differentiable ray tracer and the PSF-map rendering pipeline on which the tolerance-aware training is built.","marker":"[42]"},{"why":"Provides the differentiable engine for deep lens design that the experiments use as the base simulator.","marker":"[37]"},{"why":"Zemax is the commercial reference used to validate the accuracy of the perturbed ray tracing and as the optics-only tolerance-optimization baseline.","marker":"[43]"},{"why":"Establishes end-to-end complex lens design with differentiable ray tracing, the co-design paradigm this work extends to tolerances.","marker":"[33]"},{"why":"Closest prior deep-optics work that models a fabrication effect, DOE quantization, used as the comparison baseline in Table 1.","marker":"[22]"},{"why":"Fabrication-simulator approach from computational lithography that the paper contrasts as addressing a different optical element class.","marker":"[46]"},{"why":"Prior work that recalibrates an already fabricated design, contrasted with this paper's design-stage tolerance optimization.","marker":"[9]"},{"why":"Prior joint hardware-software calibration for manufacturing-perturbed lens systems, contrasted as post-fabrication correction rather than design-stage hardening.","marker":"[47]"}],"fun_headline_variants":["Tolerance-aware deep optics: 2 dB gain after manufacturing","Design optics that survive real-world assembly: 2 dB sharper","Closing the design-to-fabrication gap: tolerance-aware deep optics","Train optics with manufacturing errors: 2 dB gain in deblurring"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that real manufacturing and assembly errors are well described by independent Gaussian deviations of each lens with the hand-set ranges in Table A1 (for example, ±0.04 mm decentration and ±0.05 degrees tilt), since the same distribution is used for training, evaluation, and yield analysis; if a factory's actual errors are larger, correlated, or non-Gaussian, the claimed >2 dB gain may not carry over to fabricated lenses.","fun_headline_variants_meta":{"raw":{"variants":["Tolerance-aware deep optics: 2 dB gain after manufacturing","Design optics that survive real-world assembly: 2 dB sharper","Closing the design-to-fabrication gap: tolerance-aware deep optics","Train optics with manufacturing errors: 2 dB gain in deblurring"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000534,"raw_usage":{"total_tokens":2536,"prompt_tokens":884,"completion_tokens":1652,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":500,"completion_tokens_details":{"reasoning_tokens":1577}},"tokens_in":500,"tokens_out":1652,"duration_ms":14024,"temperature":1.0,"reasoning_tokens":1577,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T21:42:29.264611+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fabricate, say, 100 copies of the tolerance-optimized Lens 2, measure their actual decentration, tilt, thickness, and curvature errors, and compare the distribution of measured deblurring PSNR against the simulated 100-sample tolerance test in Table 2; if the measured performance spread matches the conventional design's spread rather than the tolerance-aware design's, the Gaussian independence model and the >2 dB claim are falsified. A cheaper preliminary check is to measure a real factory's tolerance statistics and test whether they fit the assumed normal distribution with the stated ranges.","supporting_citations":[{"cited_title":"Curriculum learning for ab initio deep learned refractive optics","cited_arxiv_id":null,"evidence_quote":"Supplies the differentiable ray tracer and the PSF-map rendering pipeline on which the tolerance-aware training is built."},{"cited_title":"do: A differ- entiable engine for deep lens design of computational imag- ing systems","cited_arxiv_id":null,"evidence_quote":"Provides the differentiable engine for deep lens design that the experiments use as the base simulator."},{"cited_title":"Opticstudio, 2013","cited_arxiv_id":null,"evidence_quote":"Zemax is the commercial reference used to validate the accuracy of the perturbed ray tracing and as the optics-only tolerance-optimization baseline."},{"cited_title":"End-to-end complex lens design with differen- tiable ray tracing","cited_arxiv_id":null,"evidence_quote":"Establishes end-to-end complex lens design with differentiable ray tracing, the co-design paradigm this work extends to tolerances."},{"cited_title":"Quantization-aware deep optics for diffractive snapshot hyperspectral imaging","cited_arxiv_id":null,"evidence_quote":"Closest prior deep-optics work that models a fabrication effect, DOE quantization, used as the comparison baseline in Table 1."},{"cited_title":"Neural Lithography: Close the Design-to-Manufacturing Gap in Computational Optics with a 'Real2Sim' Learned Photolithography Simulator","cited_arxiv_id":"2309.17343","evidence_quote":"Fabrication-simulator approach from computational lithography that the paper contrasts as addressing a different optical element class."},{"cited_title":"Computational optics for mobile terminals in mass production","cited_arxiv_id":null,"evidence_quote":"Prior work that recalibrates an already fabricated design, contrasted with this paper's design-stage tolerance optimization."},{"cited_title":"Optical degradation correction of manufacturing-perturbed glass-plastic hybrid lens systems via a joint hardware- software optimization framework","cited_arxiv_id":null,"evidence_quote":"Prior joint hardware-software calibration for manufacturing-perturbed lens systems, contrasted as post-fabrication correction rather than design-stage hardening."}],"review_version":1}