{"id":"449ac2d7-3e8d-4e6d-a2ed-8adeee49ad59","arxiv_id":"2501.00014","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Combining finite-difference PINNs with warm-started domain decomposition reproduces the lid-driven cavity corner vortices at Re=400 and Re=1000 without training on known solutions.","lead":"A team trained physics-informed neural networks in small subregions after an initial whole-domain run, recovering the small corner vortices in a standard fluid-flow benchmark without using the known answer. The approach improves accuracy near walls for Reynolds numbers 400 and 1000, making neural solvers more competitive with classical CFD on this test problem.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Corner-vortex accuracy depends on unvalidated Dirichlet data supplied from the whole-domain FD-PINN to each subdomain; the 120×120 experiment mitigates but does not close this gap.","rationale":"The reader's weakest assumption is precisely the dependence of subdomain solves on the global FD-PINN solution, and I agree it is load-bearing. The paper's own Section 5 states the limitation. The 120×120 experiment in Section 4.3 shows one failure mode is survivable, which is credit to the authors, but it is qualitative and not a quantitative validation of the boundary data. The missing measurement is easy to supply. Because the verdict CONDITIONAL already reflects these doubts, I would not change it. The strongest independent support is the direct comparison to Ghia and Botella in Table 4.1 and the 120×120 stress test; the main weakness is the single-run, no-code reporting and the absence of boundary-error quantification.","tokens_in":18637,"tokens_out":13654,"duration_ms":139176,"concrete_test":"For the Re=1000, 16-subdomain, 25×25 configuration, record the whole-domain FD-PINN velocity at all artificial subdomain boundary nodes before local training and compare with the reference trace; then rerun the subdomain stage with the reference trace substituted as the Dirichlet data, keeping all other settings fixed. If the lower-left vortex center shifts by more than about 0.01 in x or y, or wall-adjacent MSE changes by more than a factor of 2, the reported corner accuracy is controlled by unvalidated global boundary data. Repeat each run with at least three random initializations to check that the single reported run is not an optimizer accident.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Algorithm 1's subdomain stage is initialized and bounded by the whole-domain FD-PINN solution: lines 9-10 use the final trained velocity to define every artificial subdomain boundary, and Section 5 concedes that subdomain accuracy falls when the global solution is inaccurate. The central claim therefore rests on the accuracy of those boundary traces, not on the local residual alone. The paper reports MSE only on fixed lines and vortex centers, not the error of the global network at the artificial interfaces, and it does not report how the corner solution responds to perturbations of those boundary values. Section 4.3's 120×120 experiment is the right kind of stress test—the global solution lacks the lower-left vortex yet the local solve recovers it—so the concern is not automatically fatal. But that is a single qualitative example, and the 25×25 corner result is reported from a single optimization run; roughly one third of the subdomain optimizations stopped early at local minima. Without a boundary-error measurement or a boundary-perturbation test, the headline claim remains conditional.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a domain-decomposition extension of finite-difference physics-informed neural networks (FD-PINNs) for steady incompressible Navier-Stokes flow in a square lid-driven cavity. The method first trains a global FD-PINN on the whole cavity at 100×100 interior grid points, then divides the domain into 16 non-overlapping subdomains; each subdomain network is initialized with the global network weights and trained against the residual of the Navier-Stokes equations with Dirichlet velocity boundary data taken from the global solution. No inter-subdomain interface conditions are used. Results are presented for Re=400 and Re=1000, including streamline plots, vortex-center locations, and MSE tables against a 500×500 SOR reference solution, with emphasis on the lower corners where standard FD-PINNs fail to produce accurate secondary vortices. The central claim is that this two-stage procedure improves solution accuracy near the walls and generates correct secondary vortices without using the reference solution in the loss function.","tokens_in":18844,"tokens_out":8370,"duration_ms":71256,"significance":"If the results are reproducible, the paper makes a useful contribution: it shows that a simple subdomain refinement stage, initialized from a global FD-PINN, can recover corner vortices at Re=1000 where the plain FD-PINN fails, and it provides a transparent comparison against a validated reference solution (Ghia et al.). The 120×120 stress test in Section 4.3, where the global solution lacks the lower-left vortex yet the local solver recovers it, is a particularly persuasive piece of evidence. The method avoids interface conditions and is easy to implement. However, the quantitative claims are weakened by the absence of repeated runs, seeds, cost-controlled baselines, and an assessment of the sensitivity to the approximate subdomain Dirichlet data; these gaps currently make the headline accuracy improvements conditional.","major_comments":[{"comment":"The subdomain Dirichlet data are taken from the global FD-PINN solution, and Section 5 concedes that subdomain accuracy degrades if that global solution is inaccurate. The paper neither reports the error of the global solution along the artificial interfaces nor a perturbation study of these boundary values. The 120×120 experiment (Figure 12) is a single qualitative example; while it mitigates the concern, it does not quantify the dependence. Please add an interface-error table and a test in which the subdomain boundary data are perturbed or replaced by slightly different global solutions.","section":"§4.3, Algorithm 1, §5"},{"comment":"The text states that about one-third of the subdomain optimizations stopped early at local minima. No repeated runs, random seeds, or statistics are reported for the MSE tables or vortex centers. Given the small differences in Table 4.1 (e.g., x-coordinate 0.0782 vs 0.0830), the observed improvements may not be robust across initializations. Please provide mean ± standard deviation over several seeds, or a deterministic seed with a justification.","section":"§4, L-BFGS setting paragraph"},{"comment":"The accuracy comparison is not cost-controlled. The present method trains one global network plus 16 independent subdomain networks, each with up to 2×10^5 L-BFGS evaluations, whereas the standard FD-PINNs baseline is a single 100×100 training run. The claimed improvement in solution accuracy should be accompanied by wall-clock times and parameter counts, or a baseline with equivalent total compute, so that the reader can distinguish a genuine algorithmic gain from a larger training budget.","section":"§4, Tables 4.2–4.7"},{"comment":"The vortex-center locations for the present method are reported for subdomain grids 25×25, 50×50, and 100×100, but the selection of these three grid sizes is presented as a result of 'computational trials' without a criterion. Since the method's accuracy varies strongly with subdomain resolution (Tables 4.3–4.4 show different rankings for different lines), the paper should either propose a principled way to choose the subdomain resolution or report the sensitivity of the headline vortex-center claim to this choice.","section":"§4.3, Table 4.1, Figure 8"}],"minor_comments":[{"comment":"The caption contains a duplicated word: 'with with 120 × 120 grid points'.","section":"Figure 12 caption"},{"comment":"The caption reads 'lop left corner'; it should be 'top left corner'.","section":"Figure 7 caption"},{"comment":"The journal title is misspelled as 'Journal of Marchine Learning Research'; it should be 'Journal of Machine Learning Research'.","section":"Reference [17]"},{"comment":"The phrase 'we emphasizing' should be 'we emphasize'.","section":"Section 5"},{"comment":"The notation with double vertical bars and subscripts i,j is ambiguous; please clarify that the norm is evaluated at each grid point and summed over the interior points.","section":"Equation (3.2)"},{"comment":"The step 'Reshape the output variables to a matrix form and add the boundary of the velocity components' is vague; specify how the boundary conditions are imposed (hard replacement of network outputs or soft penalty) and how the finite-difference stencils treat points adjacent to the boundary.","section":"Algorithm 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the core idea is promising, but the reproducibility and cost-control issues are substantial enough that the current version should not be accepted without revision. The authors should also consider adding a data/code availability statement, as the absence of code makes it difficult to verify the early-stopping statistics and the MSE tables."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper delivers exactly what the title promises: a warm-started FD-PINN with non-overlapping domain decomposition that recovers the lower-corner vortices of the lid-driven cavity at Re=1000 and Re=400 without feeding known-solution sample points into the loss. The 120x120 stress test in Sec. 4.3, where the global solution lacks the lower-left vortex but the subdomain solve recovers it, is the clearest evidence that the mechanism is real and not just a re-plot of the reference solution. The MSE tables and vortex-center comparisons against Ghia and Botella support the claim.\n\nThe novelty is modest but legitimate. Each ingredient, FD-PINNs and DDM, is published, and warm-starting from a global run is a simple twist. The specific combination, though, removes the known-solution requirement of Jiang et al. and extends the no-known-solution range beyond the Re=400 of CAN-PINN. That is a useful step for the subfield.\n\nThe soft spots are mostly about evidence quality rather than the core idea. No code or data is released, so a careful reader cannot check the exact training details or reproduce the vortex centers. There are no repeated runs, and the paper admits roughly a third of the subdomain optimizations stopped early at local minima. Since the subdomain boundaries inherit Dirichlet data from the global FD-PINN, the accuracy of those boundary traces is load-bearing; the 120x120 experiment mitigates the concern, but the paper never reports the error of the global solution along the artificial interfaces, nor does it perturb those boundary values to see how sensitive the corner vortices are. The compute comparison to the standard FD-PINN is also not cost-controlled: the present method uses a global run plus 16 subdomain trainings, and that higher budget alone could explain part of the improvement.\n\nThe paper's own limitation statement in the conclusions is honest: it explicitly says subdomain accuracy decreases if the global solution is insufficiently accurate. That acknowledgment is to the authors' credit, and it helps define the boundary of the claim.\n\nOn balance, the central claim holds up for the specific benchmark. The weaknesses are addressable: release code and data, add repeated runs with error bars, measure or bound the artificial-interface error, and compare at matched compute. This paper deserves a serious referee, and with those additions it would be solid. For a reader working on PINNs or surrogate CFD methods, it is a worthwhile citation.\n\nRecommendation: send to peer review, with the expectation of major revisions for reproducibility and robustness.","headline":"A warm-started FD-PINN plus domain decomposition that recovers lid-driven-cavity corner vortices at Re=1000 without known-solution data; the result is plausible and useful but needs reproducibility and robustness work before it is fully convincing.","tokens_in":19396,"tokens_out":3046,"would_cite":true,"duration_ms":27149,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65N55","65M06","68T07","76D05"],"pacs":[],"model":"deepseek-v4-flash","headline":"Subdomain-refined neural nets recover corner vortices at Re=1000.","keywords":["Navier-Stokes equations","Lid-driven cavity","Finite difference methods","Physics-informed neural networks","Domain decomposition","Secondary vortices","Incompressible flow","High Reynolds number"],"falsifier":"Run the two-stage scheme starting from a deliberately degraded whole-domain solution — for example, stop the L-BFGS optimization early on the 100×100 global grid, or train that grid to a higher loss tolerance — and check whether the 16 subdomain networks still produce the lower-left secondary vortex with its benchmark center. If the vortex is recovered despite incorrect inherited boundary data, the claim that subdomain accuracy is tied to global accuracy would be contradicted; if the vortex is missed or shifted, the dependency is confirmed.","tokens_in":18414,"feed_emoji":"🌊","tokens_out":11237,"duration_ms":88856,"temperature":0.7,"pith_summary":"Physics-informed neural networks (PINNs) that rely on automatic differentiation have a hard time resolving the weak counter-rotating eddies that form in the lower corners of a lid-driven cavity at Reynolds number 1000. This paper argues that the failure lies in how derivatives are computed and in solving on one global grid, and it fixes both: replace automatic differentiation with second-order finite-difference stencils, then split the cavity into 16 subdomains and retrain a fresh network on each, seeded with the whole-domain network's weights and fed boundary values taken from that network. The paper reports that this two-stage procedure, which never uses the classical solution as training data, recovers accurate secondary vortices in both lower corners, while the standard FD-PINN on a single grid from 100×100 up to 300×300 loses the lower-left vortex. Compared with a 500×500-grid finite-difference reference solution, the refined solutions lower the mean-square error near the walls and place the vortex centers within the spread of published spectral and multigrid benchmarks. If the claim holds, it offers a simple, data-free way to push PINN solutions of the Navier–Stokes equations to higher Reynolds numbers near boundaries.","feed_headline":"Subdomain refinement lets neural nets find Re=1000 corner vortices","feed_subtitle":"Physics-informed networks retrained patch-by-patch match benchmark vortex locations without training on classical solutions.","key_machinery":"The workhorse is the FD-PINN: a physics-informed neural network whose PDE residual is evaluated not by automatic differentiation but by second-order central finite-difference stencils on a rectangular grid, with the network's outputs reshaped to the grid and boundary values inserted before differencing. The loss is the sum of squared residuals of the velocity–pressure Navier–Stokes equations and the incompressibility constraint over interior grid points, minimized with the L-BFGS optimizer. The novel part is the two-stage domain decomposition: a whole-domain FD-PINN is trained at 100×100 points, and then the domain is cut into 16 rectangular subdomains, each fitted with its own network initialized from the whole-domain weights and biases; each subdomain's boundary conditions are extracted from the already-trained global solution, which removes any need for interface conditions between neighboring subdomains and lets the local networks refine structure (such as corner eddies) that the global network resolved poorly.","core_discovery":"The authors claim that the steady incompressible Navier–Stokes equations in a lid-driven cavity can be solved accurately at Re=1000 by an FD-PINN without any known-solution data, provided the problem is split across scales: first train one network on the whole domain with a moderate 100×100 grid, then divide the cavity into 16 rectangular subdomains and retrain an independent network in each, initialized from the whole-domain weights and biases and constrained on its boundaries by the whole-domain network's output. The loss in both stages is the squared PDE residual (momentum equations plus the divergence-free constraint) evaluated with central finite differences on the grid, so no interface conditions between subdomains are needed. On this scheme, the lower-left secondary vortex, which standard FD-PINNs produce only faintly at 100×100 and lose entirely at finer grids, appears clearly at all three subdomain grid resolutions, and the most accurate vortex centers come from the coarsest 25×25 subdomain grid. The authors quantify the gain with mean-square errors against a finite-difference reference solution, showing substantial error reduction in the wall-adjacent strips, and they demonstrate that the same refinement rescues the vanished lower-left corner vortex at Re=400. The central assertion is that the dominant error source in FD-PINNs is the overprediction that comes from training a single network on a fine global grid, and that local retraining on coarse grids eliminates that overprediction while preserving accuracy in the interior.","pith_inferences":["The pattern of results — coarse grids winning in the low-intensity lower corners and finer grids winning near the lid — suggests an automatic adaptive strategy: let the network choose its own local grid density based on velocity magnitude or gradient, which could cut training cost below the fixed 16×16 split.","The two-stage scheme resembles a classical defect-correction or multigrid idea, with the whole-domain network acting as a coarse solver and subdomain retraining as local smoothing; if the analogy holds, iterating the global-refine cycle (re-train the global network from the refined subdomains) might push the method to still higher Reynolds numbers.","A caution implied by the method: because subdomain boundaries come from the global network without any interface constraint, a wrong global solution can be faithfully reproduced rather than corrected; a testable variant would add overlap or interface consistency to see whether the vortex centers shift toward the benchmarks.","The reference solution is itself numerical, not experimental; the next discriminating test for the claim that FD-PINNs resolve true corner physics would be comparison against physical experiments or a spectral solver at the same Reynolds number."],"forward_implications":["At Re=1000, the subdomain-refined FD-PINN produces a clear secondary vortex in the lower-left corner at every subdomain grid tested (25×25, 50×50, 100×100), whereas standard FD-PINNs on 100×100, 120×120, 150×150, and 300×300 grids lose or distort that vortex.","The 25×25 subdomain grid gives the vortex-center locations closest to the reference values (x ≈ 0.0735, y ≈ 0.0753 for the lower left; x ≈ 0.8695, y ≈ 0.1210 for the lower right).","Along the horizontal line y=0.99, the mean-square error of the vertical velocity component drops from about 2.6×10^-3 for the standard FD-PINN to about 1.2×10^-5 for the present method with 100×100 subdomain grids, a gain of roughly two orders of magnitude near the moving lid.","The same refinement scheme restores the completely vanished lower-left corner vortex at Re=400 using only 25×25 subdomain grid points, so the benefit is not specific to Re=1000.","Because subdomains are trained independently with no interface conditions, the refinement step is embarrassingly parallel, so the approach can be distributed across cores without changing the algorithm."],"supporting_citations":[{"why":"The original finite-difference PINN study, which showed that without known-solution sample points the method misses secondary vortices; it is the baseline that the present technique improves on.","marker":"[23]"},{"why":"A finite-difference Navier-Stokes solver with SOR pressure iteration; the paper's reference solution is generated with this scheme on a 500×500 grid.","marker":"[38]"},{"why":"The classical multigrid benchmark for lid-driven cavity flow that supplies the reference against which the paper's velocity profiles, vortex locations, and overall accuracy are judged.","marker":"[39]"},{"why":"A spectral benchmark for the lid-driven cavity; its reported vortex-center values appear in the paper's comparison Table 4.1.","marker":"[41]"},{"why":"A PINN variant using coupled automatic and numerical differentiation that solved Re=400 without known solutions; it is the prior accuracy benchmark for data-free PINN approaches.","marker":"[24]"},{"why":"A domain-decomposition PINN approach for the unsteady lid-driven cavity, the strategy the paper adapts to steady FD-PINNs.","marker":"[31]"}],"fun_headline_variants":["Subdomain splits let FD-PINNs pinpoint Re=1000 corner vortices","Retraining on coarse subgrids reclaims lost Re=1000 corner vortices","FD-PINN domain splitting recovers secondary vortices without data","Patchwise retraining sharpens FD-PINN corner vortex positions","Coarse grid patches boost FD-PINN accuracy for Re=1000 cavity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method rests on the accuracy of the boundary values that the whole-domain FD-PINN supplies to each subdomain; if that global solution is wrong in a corner, the local network inherits the error and cannot recover the true vortex on its own.","fun_headline_variants_meta":{"raw":{"variants":["Subdomain splits let FD-PINNs pinpoint Re=1000 corner vortices","Retraining on coarse subgrids reclaims lost Re=1000 corner vortices","FD-PINN domain splitting recovers secondary vortices without data","Patchwise retraining sharpens FD-PINN corner vortex positions","Coarse grid patches boost FD-PINN accuracy for Re=1000 cavity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000855,"raw_usage":{"total_tokens":3747,"prompt_tokens":1014,"completion_tokens":2733,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":630,"completion_tokens_details":{"reasoning_tokens":2632}},"tokens_in":630,"tokens_out":2733,"duration_ms":15623,"temperature":1.0,"reasoning_tokens":2632,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T15:34:19.177670+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the two-stage scheme starting from a deliberately degraded whole-domain solution — for example, stop the L-BFGS optimization early on the 100×100 global grid, or train that grid to a higher loss tolerance — and check whether the 16 subdomain networks still produce the lower-left secondary vortex with its benchmark center. If the vortex is recovered despite incorrect inherited boundary data, the claim that subdomain accuracy is tied to global accuracy would be contradicted; if the vortex is missed or shifted, the dependency is confirmed.","supporting_citations":[{"cited_title":"Applications of finite difference-based physics-informed neural networks to steady incompressible isothermal and thermal flows,","cited_arxiv_id":null,"evidence_quote":"The original finite-difference PINN study, which showed that without known-solution sample points the method misses secondary vortices; it is the baseline that the present technique improves on."},{"cited_title":"Griebel, T","cited_arxiv_id":null,"evidence_quote":"A finite-difference Navier-Stokes solver with SOR pressure iteration; the paper's reference solution is generated with this scheme on a 500×500 grid."},{"cited_title":"High-Re solutions for incompressible flow using the Navier-Stokes equations and a multigrid method,","cited_arxiv_id":null,"evidence_quote":"The classical multigrid benchmark for lid-driven cavity flow that supplies the reference against which the paper's velocity profiles, vortex locations, and overall accuracy are judged."},{"cited_title":"Benchmark spectral results on the lid-driven cavity flow,","cited_arxiv_id":null,"evidence_quote":"A spectral benchmark for the lid-driven cavity; its reported vortex-center values appear in the paper's comparison Table 4.1."},{"cited_title":"CAN-PINN: A fast physics-informed neural network based on coupled-automatic–numerical differentiation method,","cited_arxiv_id":null,"evidence_quote":"A PINN variant using coupled automatic and numerical differentiation that solved Re=400 without known solutions; it is the prior accuracy benchmark for data-free PINN approaches."},{"cited_title":"Physics-informed neural networks with domain decomposition for the incompressible Navier–Stokes equations,","cited_arxiv_id":null,"evidence_quote":"A domain-decomposition PINN approach for the unsteady lid-driven cavity, the strategy the paper adapts to steady FD-PINNs."}],"review_version":1}