{"id":"19b20896-8fa4-439e-9d74-19a62e321623","arxiv_id":"2509.03622","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A neural subdomain preconditioner plus multilevel domain decomposition solves 2D Maxwell problems up to 200 wavelengths and drives inverse design of large nanophotonic devices.","lead":"The authors combine a neural-network error predictor with classical iterative solvers to simulate 2D Maxwell's equations on large, variable-size domains, and they use the same solver to inverse-design multi-wavelength photonic devices. The significance is a potential pathway to accurate neural surrogates for real nanophotonics, although the method is currently slower than direct solvers on the tested 2D problems.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Tables 1–2 establish near-optimal scaling only for smoothly varying grayscale dielectrics, while §2.2 concedes that discrete material boundaries and resonant subdomains degrade the coarse space—exactly the binary regime used in Figure 6 inverse design. The 200-wavelength scalability claim is therefo","rationale":"This is the same load-bearing concern identified by the reader, and the paper's own language supports it: Tables 1–2 use only smooth grayscale media, while §2.2 explicitly warns that discrete material boundaries and resonant subdomains worsen coarse-space effectiveness, and the Discussion reports sensitivity to overlap/momentum and slow coarse-space degradation with problem size. The abstract promises solving problems with 'different scales, resolutions, and dielectric media distribution,' so the unmeasured binary/high-contrast regime is not a peripheral limitation but a direct gap in the central scalability claim. I do not see a mathematical inconsistency in the iterative scheme itself; the subdomain preconditioner and inverse-design comparisons are substantial. The issue is evidentiary coverage: near-optimal two-level scaling has been demonstrated only for the most favorable dielectric class. Because the reader's conditional verdict already accounts for this and asks for additional data and code, I do not move the verdict. A targeted experiment on binary media, using the same protocol as Table 2, would settle whether the concern lands.","tokens_in":15660,"tokens_out":5152,"duration_ms":60792,"concrete_test":"Reproduce Table 2's two-level scaling with the 256×256 subdomain model on random binary Voronoi/freeform dielectric maps with ε∈{1,8} and feature sizes 1–50 px (the stated training distribution), for global grids (994, 1486, 1978, 2470, 2962), fixing kcoarse and overlap as in Table 2 and Algorithm 3. Record 1L/2L iteration counts and relative residual/L1 error. If the 2L iteration counts are no longer nearly flat (e.g., more than ~2× growth per doubling of subdomain count, or kcoarse must grow with problem size to keep iterations bounded), then the 200λ scalability claim fails for binary media. Also repeat a few runs with ±10% overlap and momentum perturbations to test the acknowledged sensitivity.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that a single trained subdomain neural operator, used as an F-GMRES preconditioner and as an engine for coarse-space eigenproblems, yields a two-level overlapping Schwarz solver that is accurate and scalable up to ~3000×3000 grids / 200 wavelengths, and enables inverse design >100λ matching FDFD. The load-bearing condition is that the NN-constructed coarse space remains effective for the dielectric distributions the paper advertises. That condition is only tested in Tables 1 and 2 for media that are, in the authors' words, 'grayscale and varies smoothly'; the paper immediately adds that 'highly resonant subdomains with discrete material boundaries worsen the effectiveness of the coarse space.' This caveat covers the binary, high-contrast, sharply structured devices in Figure 6 (the multiplexer and metalens are explicitly binary), so the scaling evidence does not extend to the inverse-design regime or to general dielectric maps. The Discussion repeats that 'the effectiveness of the coarse space also slowly degraded as the problem size increased' and that convergence was increasingly sensitive to overlap and momentum. Since the tables report iteration counts with no error bars, no kcoarse selection rule, and no hyperparameter-robustness data, the near-optimal scaling claim is not presently falsifiable from the manuscript. The inverse-design comparison with an FDFD ground truth is real evidence that the solver can produce useful engineering optima, but it does not establish scaling for binary media.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a hybrid neural-operator/iterative solver for 2D frequency-domain Maxwell equations. A Fourier neural operator is trained to map residuals to error fields and is used as a flexible preconditioner inside F-GMRES; the same model is used to build approximate Robin-to-Robin maps for coarse-space construction in a two-level overlapping Schwarz method. The authors claim machine-precision global solutions, near-optimal iteration scaling up to roughly 3000x3000 grids (about 200 wavelengths), and inverse design of devices larger than 100 wavelengths that matches an FDFD ground truth. The manuscript reports one-level Schwarz results (Figure 4), scaling tables for 64x64 and 256x256 subdomain models (Tables 1-2), and three inverse-design demonstrations (Figure 6).","tokens_in":16000,"tokens_out":7316,"duration_ms":79302,"significance":"The central design choice is sound: convergence is checked against the true FDFD operator residual, so the neural network acts as a preconditioner rather than a standalone approximate solver. This is a meaningful step toward using neural surrogates in a way that preserves a user-controlled accuracy guarantee. The recurrent F-GMRES training scheme, the modulation-based FNO architecture, and the use of the subdomain model for coarse-space eigenproblems are interesting contributions, and the ground-truth FDFD comparison in the inverse-design section is a genuine strength. However, the headline claims outrun the evidence: the 'machine precision' statement is unsupported by the reported losses, and the scaling results are restricted to smoothly varying grayscale media, while the paper itself reports degradation for discrete-material/resonant subdomains and increasing sensitivity to hyperparameters as the problem grows. The significance is therefore conditional on strengthening and honestly scoping the empirical claims.","major_comments":[{"comment":"The claim that the subdomain preconditioner 'can combine with one-level overlapping Schwarz methods to solve global problems with machine precision' is not supported by the data. Figure 4 reports L1 losses down to 3.6e-6 in the best case and 9.9e-5 in others, with final-error plots drawn on a ±1% scale and simulations in single precision. These are residual-bounded accuracies, not machine precision. Since this accuracy statement is central to the paper's promise, either replace 'machine precision' with 'prescribed residual accuracy' or provide evidence of convergence to roughly 1e-7 in the reported metric.","section":"§2.2, Figure 4"},{"comment":"The near-optimal scaling claim is established only for dielectric maps that the authors themselves describe as 'grayscale and varies smoothly.' The same paragraph concedes that 'highly resonant subdomains with discrete material boundaries worsen the effectiveness of the coarse space,' and the Discussion adds that convergence became increasingly sensitive to overlap and momentum and that coarse-space effectiveness 'slowly degraded' as problem size increased. The inverse-design devices in Figure 6 are explicitly binary, so the 200-wavelength scalability claim is not supported in the regime the paper advertises. Moreover, Tables 1-2 do not state the global residual tolerance, number of random realizations, error bars, or a rule for choosing kcoarse; these omissions make the iteration counts difficult to interpret. Please add scaling experiments on binary/high-contrast media and report conve","section":"§2.2, Tables 1-2 and §3"},{"comment":"The inverse-design evidence for 'accurate' large-scale optimization is limited. As written, the ground-truth FDFD comparison is explicitly made for the multiplexer optimization trajectory; the coupler and metalens are presented without an independent FDFD comparison, even though these are the binary/large-scale cases where the coarse space is most at risk. If the ground-truth comparison is meant to cover all three devices, the manuscript should state this and show the quantitative convergence/efficiency comparisons. Otherwise, a second ground-truth comparison is needed to support the generalization claimed in the text.","section":"§2.2, Figure 6"}],"minor_comments":[{"comment":"Typo in the table title: 'Schwawrz' should be 'Schwarz.' In the Discussion, 'course grid construction' should be 'coarse grid construction.'","section":"Table 2"},{"comment":"The final-error plots are described as spanning -1% to 1% of the maximum ground-truth field, but no colorbar or normalization formula is provided. Please add a colorbar and state whether the L1 losses are normalized by the L1 norm of the ground truth.","section":"Figure 4"},{"comment":"The training recurrence N, clipping threshold th_clip, and residual thresholds are load-bearing hyperparameters, but no default values or sensitivity analysis are reported. Please provide the values used in the experiments for reproducibility.","section":"Algorithms 1-2"},{"comment":"Training details for the 256x256 physics-only model are sparse, and the manuscript repeatedly refers to the Supplementary Information for details. If the arXiv version is the review version, include the SI or summarize the omitted training hyperparameters and schedules.","section":"§4.1"},{"comment":"The Discussion notes that the solver is about 2x slower than PARDISO for the tested problems and would only become advantageous for much larger problems. This is important context for the scalability claims and should be mentioned in the abstract or introduction as well.","section":"§3"}],"recommendation":"major_revision","confidential_remarks":"The core computational idea is sound, and the manuscript is likely fixable with honest scoping and additional experiments. The main risk is claim-evidence mismatch: 'machine precision' and '200-wavelength scalability' are not supported by the current evidence, and the authors themselves acknowledge degradation in the binary/resonant regime. I would not reject, because the true-residual checking and the FDFD ground-truth inverse-design benchmark are genuine strengths."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis paper deserves a serious look. The core idea is genuinely new: a single neural operator trained on subdomain problems with Robin boundary conditions, used both as an F-GMRES preconditioner and as a fast evaluator of Robin-to-Robin maps to build spectral coarse spaces for two-level overlapping Schwarz. That integration is not a mashup; it addresses a real bottleneck in domain decomposition for indefinite Maxwell systems. The inverse design comparison against an FDFD ground truth is convincing evidence that the solver can produce useful optima, and the authors are refreshingly explicit about limitations—e.g., that the coarse space degrades for resonant, discrete material boundaries and that the speed advantage over PARDISO is projected, not measured, at scale.\n\nWhere the paper overreaches: the headline \"machine precision\" is not supported by the L1 errors in Figure 4 (1e-5 at best), and \"accelerated\" is a stretch given the PARDISO benchmark showing the GPU solver is 2x slower. The two-level scaling tables use only smoothly varying grayscale media; the paper itself concedes that discrete, high-contrast boundaries—exactly what appears in the binary inverse-design devices—worsen the coarse space. So the 200-wavelength scalability claim does not yet extend to the advertised general dielectric distributions. Also, no code or data, and the SI is missing key hyperparameters, so replication is currently impossible. These are addressable with a code release, more test cases with binary media, and tightened claims.\n\nWho should read it: people working on neural preconditioners, domain decomposition, and computational photonics. It deserves peer review—conditional accept at strongest—but the authors should be asked for code, a robustness study on binary media, and an honest reframing of \"machine precision\" and \"accelerated.\" I'd bring it to reading group and likely cite it if they release the code.","headline":"A genuinely new integration of neural preconditioners with spectral coarse spaces for two-level Schwarz, worth a serious referee despite some overclaimed headlining.","tokens_in":64,"tokens_out":1373,"would_cite":true,"duration_ms":44131,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A neural operator trained on subdomains solves Maxwell problems on grids up to 3000 by 3000.","keywords":["neural operator","Maxwell's equations","domain decomposition","overlapping Schwarz","coarse space","F-GMRES preconditioning","inverse design","FDFD"],"falsifier":"Run the same 256x256 subdomain network and two-level Schwarz method on a global domain built from binary random blocks of permittivity 1 and 8 with sharp edges, at 4x4, 8x8, and 12x12 subdomains with a fixed kcoarse, and compare the number of iterations to the grayscale cases reported in Table 2. If the iteration counts grow with the number of subdomains at a rate close to the one-level method, the near-optimal scaling claim is falsified for that material class.","tokens_in":15510,"feed_emoji":"⚡","tokens_out":8644,"duration_ms":78854,"temperature":0.7,"pith_summary":"The paper argues that neural PDE surrogates can be both accurate and scalable if they are embedded in iterative multilevel solvers instead of being asked to produce global solutions directly. It trains a single subdomain neural operator on small 64x64 or 256x256 boundary-value problems with arbitrary Robin-type boundary conditions, then reuses that network in two ways: as a preconditioner inside an F-GMRES loop to solve each subdomain to a chosen residual tolerance, and as an engine for building the coarse space of a two-level overlapping Schwarz method. With this setup, the authors report solving 2D frequency-domain Maxwell problems on global grids up to about 3000 by 3000 and physical sizes up to about 200 wavelengths, with iteration counts that stay nearly constant on smoothly varying dielectric distributions. The same solver drives adjoint-based inverse design of photonic devices larger than 100 wavelengths, producing layouts and efficiencies nearly identical to a ground-truth FDFD solver. If the claims hold, neural surrogates gain a credible path to large, practical electromagnetic simulation with controllable accuracy.","feed_headline":"One neural operator scales Maxwell solves to 200 wavelengths","feed_subtitle":"Reused as preconditioner and coarse-space engine, one subdomain model makes large photonic simulations accurate.","key_machinery":"The central object is a modified Fourier neural operator that maps a residual field, together with auxiliary permittivity, source, and PML maps, to an approximate error field. It is trained inside a recurrent F-GMRES loop, so the network sees the same residual-error distribution it will encounter at inference. Its two load-bearing uses are (1) as a flexible preconditioner generating Krylov directions for inner subdomain solves, and (2) as a fast evaluator of Robin-to-Robin maps: pseudo-plane-wave probe boundary values are fed through the network, neighboring boundary values are extracted, and Rayleigh-Ritz on this compact linear map produces the coarse-space eigenvectors. These eigenvectors","core_discovery":"In its own terms, the paper's central claim is that a subdomain neural operator can be turned into an accurate global Maxwell solver by nesting two iterative loops: an inner network-preconditioned F-GMRES that solves subdomain problems with bounded residual, and an outer overlapping Schwarz iteration that updates subdomain boundary conditions until the global residual is small. The same network accelerates coarse-space construction: probe Robin boundary conditions are pushed through the network, the resulting Robin-to-Robin map is assembled, and Rayleigh-Ritz on that map yields slowly decaying modes used as global basis functions. The paper shows one-level overlapping Schwarz can reach machi","pith_inferences":["The scaling tables use smoothly varying grayscale permittivity; the paper's own observation that resonant subdomains with discrete material boundaries degrade the coarse space implies the near-constant iteration counts should be re-tested on binary, high-contrast, sharply edged layouts before claiming general nanophotonic scalability.","The network's implicit learning of wavelength and resolution from the residual encoding suggests the same recurrent-training recipe may carry over to other parameter-dependent PDE families, such as Helmholtz or elastodynamics, without explicit frequency inputs; the paper does not claim this transfer.","If the coarse space is the bottleneck, then a more direct test of the method's ceiling is how many coarse modes kcoarse must grow as subdomains become more resonant; measuring that growth rate would separate network-precision limits from coarse-space limits.","The reported speed comparison is at 2D sizes where direct sparse solvers still fit in memory; the decisive regime the paper anticipates is larger 2D and 3D problems, so a battery of memory-bound 3D Maxwell benchmarks would test whether the batching and coarse-space setup retains its advantage."],"forward_implications":["One trained subdomain model transfers across global domain sizes, resolutions, wavelengths, and dielectric maps, because the network only ever receives subdomain-sized inputs.","Global accuracy is set by residual thresholds rather than by the network's fixed approximation error, giving users a direct accuracy-versus-cost dial.","The coarse space is built once per dielectric layout and then reused for many source configurations, which is the case that arises in optimization and design sweeps.","Because training and inference share the same F-GMRES loop, the network is optimized for how it is actually deployed, and the training recurrence itself expands the effective dataset from residual-error pairs produced along the iteration path."],"supporting_citations":[{"why":"Supplies the Fourier neural operator architecture the paper modifies with a modulation path and low-rank spectral weights.","marker":"[15]"},{"why":"Prior iterative neural domain-decomposition solver for subdomain problems; the paper extends it to accurate F-GMRES preconditioning with arbitrary Robin conditions.","marker":"[24]"},{"why":"Provides the efficient iterative solvers used as baselines in the 20-50x iteration-count comparison.","marker":"[35]"},{"why":"Supplies the domain-decomposition framework and the motivation for coarse-space corrections in two-level Schwarz methods.","marker":"[36]"},{"why":"Establishes spectral coarse spaces from local Dirichlet-to-Neumann maps, the template for the paper's Robin-to-Robin eigenproblem.","marker":"[37]"},{"why":"Provides the robust coarse-space construction for optimized Schwarz methods that the paper adapts to neural operator-evaluated maps.","marker":"[38]"},{"why":"Shows near-optimal two-level domain-decomposition scaling for high-frequency Helmholtz using plane-wave coarse spaces, the theoretical basis for interpreting kcoarse as extra global iterations.","marker":"[39]"},{"why":"Supplies the FDFD solver used to generate the ground-truth subdomain training data for the 64x64 model.","marker":"[41]"}],"fun_headline_variants":["Multilevel neural operator scales Maxwell's equations to 200 wavelengths","Neural operator preconditioner powers accurate Maxwell simulation at scale","Nested iterative loops with neural operator for large Maxwell solves","Subdomain neural operator accelerates Maxwell simulation up to 200 wavelengths"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The claim that iteration counts stay almost constant as the domain grows rests on coarse spaces that work well for smoothly varying grayscale dielectrics; if the coarse space loses effectiveness on sharp, high-contrast, resonant material distributions, the advertised 200-wavelength scalability does not carry over.","fun_headline_variants_meta":{"raw":{"variants":["Multilevel neural operator scales Maxwell's equations to 200 wavelengths","Neural operator preconditioner powers accurate Maxwell simulation at scale","Nested iterative loops with neural operator for large Maxwell solves","Subdomain neural operator accelerates Maxwell simulation up to 200 wavelengths"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000917,"raw_usage":{"total_tokens":3751,"prompt_tokens":702,"completion_tokens":3049,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":446,"completion_tokens_details":{"reasoning_tokens":2992}},"tokens_in":446,"tokens_out":3049,"duration_ms":25128,"temperature":1.0,"reasoning_tokens":2992,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T10:47:58.349591+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same 256x256 subdomain network and two-level Schwarz method on a global domain built from binary random blocks of permittivity 1 and 8 with sharp edges, at 4x4, 8x8, and 12x12 subdomains with a fixed kcoarse, and compare the number of iterations to the grayscale cases reported in Table 2. If the iteration counts grow with the number of subdomains at a rate close to the one-level method, the near-optimal scaling claim is falsified for that material class.","supporting_citations":[{"cited_title":"In: Salakhutdinov, R., Kolter, Z., Heller, K., Weller, A., Oliver, N., Scarlett, J., Berkenkamp, F","cited_arxiv_id":null,"evidence_quote":"Prior iterative neural domain-decomposition solver for subdomain problems; the paper extends it to accurate F-GMRES preconditioning with arbitrary Robin conditions."},{"cited_title":"Advances in Water Resources 34(9), 1124–1139 (2011)","cited_arxiv_id":null,"evidence_quote":"Provides the efficient iterative solvers used as baselines in the 20-50x iteration-count comparison."},{"cited_title":"SIAM, ??? (2015)","cited_arxiv_id":null,"evidence_quote":"Supplies the domain-decomposition framework and the motivation for coarse-space corrections in two-level Schwarz methods."},{"cited_title":"Comptes Rendus","cited_arxiv_id":null,"evidence_quote":"Establishes spectral coarse spaces from local Dirichlet-to-Neumann maps, the template for the paper's Robin-to-Robin eigenproblem."},{"cited_title":"Comptes Rendus Mathematique 353(10), 959–963 (2015)","cited_arxiv_id":null,"evidence_quote":"Provides the robust coarse-space construction for optimized Schwarz methods that the paper adapts to neural operator-evaluated maps."},{"cited_title":"Numerische Mathematik 85(2), 283–308 (2000)","cited_arxiv_id":null,"evidence_quote":"Shows near-optimal two-level domain-decomposition scaling for high-frequency Helmholtz using plane-wave coarse spaces, the theoretical basis for interpreting kcoarse as extra global iterations."},{"cited_title":"ACS Photonics 6(11), 3010–3016 (2019) 23","cited_arxiv_id":null,"evidence_quote":"Supplies the FDFD solver used to generate the ground-truth subdomain training data for the 64x64 model."}],"review_version":1}