{"id":"e3a1bd33-4edb-4dfb-a4cd-26922dd8708d","arxiv_id":"2607.06237","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A per-instance learned graph-network transfer between coarse and fine discrete-field optimizations improves Darcy and EIT coefficient reconstructions without surrogates or pretraining.","lead":"The paper introduces a multilevel method for PDE inverse problems that optimizes discrete fields level-by-level and learns a graph-based transfer operator between grids, without offline training. On a manufactured Darcy test it reports much lower error at moderately more compute, and on public EIT data it slightly beats the official linearized reconstruction.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Abstract's '1.25x cost' omits the 3000-step ResiPhy-GNN transfer fitting at the 128^2 target; including that cost (Eq. 38) raises the ratio to ~1.75x, weakening the 'similar cost' claim.","rationale":"The most load-bearing concern is the completeness of the cost metric used in the headline efficiency claim. The abstract's '1.25x' is the numerical anchor for 'at similar cost.' Eq. (38) explicitly defines C_grid as the sum of DNF solver steps across levels, and Section 5.1 reveals that each transfer requires 3000 additional Adam steps on the target grid. These steps are not counted. A first-order correction gives ~1.75x, a 40% undercount. This matters because the central claim is not merely that the multilevel initialization improves accuracy (which the matched LUNDIsim ablation supports), but that this improvement comes at near-equal cost. If the true cost ratio is 1.75x, the wording 'similar cost' is arguable; if it is higher after counting all overhead (GNN forward/backward, feature computation), the quantitative claim may need substantial revision. The fix is straightforward and does not require new experiments: report the corrected proxy or explicitly state that the proxy excludes transfer fitting. The EIT comparison is also confounded by optimizer/regularization changes, but the paper already acknowledges that the causal transfer claim rests on the matched LUNDIsim experiment, so that confound is secondary. The reader's CONDITIONAL verdict already identifies this cost-accounting issue and conditions acceptance on reporting it; my stress-test agrees, so the verdict should remain CONDITIONAL (UNCHANGED).","tokens_in":24142,"tokens_out":7591,"duration_ms":69849,"concrete_test":"Recompute C_grid in Eq. (38) for the Darcy 64^2->128^2 run, adding for each interface the cost of the transfer-fitting steps: for each of 3000 Adam steps, add the target grid size n_{l+1}^2 (with the same per-step weighting as a DNF step). For the four-level LUNDIsim run, add the three transfer fitting costs at their respective target grids (32^2, 64^2, 128^2). Report the corrected ratio C_grid^corr/C_direct. If the corrected ratio is >=1.6, the abstract's '1.25x' claim is inaccurate and should be revised; if the corrected ratio is still <1.5, the efficiency claim largely stands. Ideally, also report wall-clock times to validate the proxy.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest quantitative evidence for the central claim is the controlled Darcy comparison: the 64^2->128^2 path is reported to use 1.25x the cumulative grid-work proxy of the direct 128^2 solve, defined in Eq. (38) as C_grid = sum_k s_k n_k^2. This proxy counts only the DNF Adam steps at each level (6000 at 64^2 and 6000 at 128^2). It omits the per-instance fitting of ResiPhy-GNN, which Section 5.1 states is '3000 Adam steps' and which is evaluated at the target 128^2 grid (the transfer loss (29) uses the refined PDE residual and observations on the target grid). Including these 3000 steps at 128^2 increases the numerator from 1.25 to approximately 1.75 (6000*4096 + 6000*16384 + 3000*16384 = 172M vs 98M for direct). The abstract's '1.25x' therefore understates the full method's cost by about 40% (relative). The four-level LUNDIsim run (Sec. 5.4) has three transfers, each 3000 steps, so the undercount is larger. Since the central claim is 'better reconstructions than direct single-level DNF optimization at similar cost,' the quantitative cost basis is incomplete. The qualitative conclusion (large error reduction for modest extra cost) may survive correction, but the headline number and the 'similar cost' phrasing should be revised.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces ResiPhy-MDNF, a multilevel framework for PDE-constrained coefficient inverse problems. On each resolution level, a Discrete Neural Field (DNF) optimizer directly updates discrete state and coefficient arrays under the prescribed numerical model, without surrogate models or offline training. Between levels, a per-instance graph neural network, ResiPhy-GNN, is fitted to the target-level transfer loss and produces a corrected coarse-to-fine initialization. The method is instantiated for full-space Darcy permeability inversion and for reduced linearized complete-electrode-model EIT on the KTC2023 dataset. The main reported results are: a controlled Darcy comparison in which the 64^2→128^2 path uses 1.25× the cumulative grid-work proxy of the direct 128^2 solve while reducing coefficient and state errors by factors of 6.76 and 10.4; a matched LUNDIsim ablation in which ResiPhy-GNN transfer improves final errors by about 8–9% over simple interpolation; and a KTC2023 EIT pipeline that reaches 0.623 mean mIoU versus 0.603 for the official linearized CEM reconstruction and 0.533 for direct single-level DNF. The paper argues that the same multilevel construction transfers across different PDEs, discretizations, and observation geometries.","tokens_in":24528,"tokens_out":6310,"duration_ms":65029,"significance":"If the results hold, the framework is a practical and interesting contribution: it retains the prescribed numerical operator, requires no offline training, and shows that a per-instance learned transfer can improve coarse-to-fine inverse optimization. The matched LUNDIsim ablation in Section 5.3 is a genuine strength: it isolates the transfer mechanism from the multilevel path and shows a consistent improvement under identical fine-grid optimization budgets. The per-instance fitting of the transfer also avoids the circularity risk that would come from using fine-grid truth during transfer training. The main quantitative claim, however, is the efficiency-precision trade-off in the Darcy experiment, and that claim rests on a cost proxy that omits a substantial part of the actual computation. The qualitative conclusion may survive correction, but the headline numbers and the wording need revision. The KTC2023 comparison is less controlled than the Darcy/LUNDIsim comparisons and should be presented with appropriate caveats or supplemented with a matched baseline.","major_comments":[{"comment":"The headline '1.25× cumulative grid-work proxy' omits the cost of fitting ResiPhy-GNN. Section 5.1 states that correctors are fitted for 3000 Adam steps, and the Darcy transfer loss (29) is evaluated on the refined target grid. Equation (38), however, counts only the DNF-level steps s_k=6000 at each grid. For the 64^2→128^2 basic case, the reported numerator is 6000·64^2 + 6000·128^2, giving 1.25× the direct 128^2 cost. Adding the 3000 transfer-fitting Adam steps at 128^2 gives 6000·64^2 + 6000·128^2 + 3000·128^2, which is about 1.75× the direct cost, not 1.25×. For the four-level LUNDIsim run, three transfer fittings are omitted, so the undercount is larger. The manuscript is internally consistent about what Eq. (38) counts, but Section 5.2 calls this a 'quantified total-work comparison,' and the abstract presents 1.25× without qualification. This is load-bearing for the 'similar cost'","section":"§5.2, Eq. (38); §5.1; Abstract"},{"comment":"The KTC2023 comparison between single-level DNF and ResiPhy-MDNF is not matched. The single-level baseline uses 10,000 Adam updates with the official-setting regularization, while the complete pipeline uses 1200 coarse-grid Adam updates, eight corrector L-BFGS iterations, and 60 outer mesh-L-BFGS updates, together with the full measurement matrix and the adaptively chosen conductivity weight (34). The text acknowledges that 'an Adam and an L-BFGS update differ in cost,' but the table and the introduction still present the mIoU gain as evidence in favor of the complete pipeline. Because the optimizer, loss weighting, and computational budget all differ, this comparison does not isolate the multilevel/DNF effect. The matched LUNDIsim ablation is the correct evidence for the transfer, and the KTC2023 section should be framed as a full-pipeline comparison only. If the cross-physics claim is","section":"§5.5, Table 5"},{"comment":"The claim that the framework is 'the same multilevel construction' across Darcy and EIT is somewhat overstated because the EIT realization has only one transfer interface and two DNF stages, and it uses a linearized reduced-space model rather than the full-space DNF formulation used for Darcy. This is not an error, but the paper's framing in the abstract and introduction suggests a more uniform hierarchy than is actually instantiated. The transfer-loss and feature constructions are analogous, and the matched LUNDIsim experiment provides the controlled evidence, so this is a presentation issue rather than a technical flaw. I recommend that the cross-physics claim be stated as 'the same transfer module can be instantiated for two different linear/full-space formulations,' which is well supported.","section":"§2.2, §3.3, Table 3"}],"minor_comments":[{"comment":"The 'relative L2 errors' E_K and E_U used in Table 2 and Eq. (39) are not formally defined. Please state whether these are normalized by the L2 norm of the reference field and over which grid points or cells they are computed.","section":"§5.1, §5.2"},{"comment":"The transfer ablation in Table 3 is reported for a single LUNDIsim instance without multiple random seeds or repeated runs. The reported improvements of about 8–9% are modest; please state whether the optimizations are deterministic and, if not, report mean and variance over seeds.","section":"§5.3, Figure 4"},{"comment":"The measurement-energy rule for lambda_sigma is calibrated once from Data 1 and then applied to all phantoms. The paper states this clearly, but please make explicit in the text that the calibration does not see the reconstructions and is not tuned per phantom, so that the reader does not confuse this with an oracle choice.","section":"§4.2, Eq. (34)"},{"comment":"There are numerous missing spaces in the abstract and main text around math and words (e.g., '642 →128 2', 'medium' appears as 'medum', etc.). These appear to be PDF-extraction artifacts, but the final journal version should have a clean text layer.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The core idea is sound and the matched LUNDIsim ablation is the strongest part of the evidence. The main obstacle is the efficiency claim in the abstract and Section 5.2: the 1.25× number omits the 3000-step transfer-fitting cost, and the phrase 'total-work comparison' overstates what Eq. (38) measures. This is fixable by changing the accounting or the wording, but it is load-bearing for the paper's headline contribution. The KTC2023 section would also benefit from a matched single-level baseline or an explicit wall-clock/FLOPs comparison, otherwise the cross-physics claim rests on a pipeline-level comparison with different optimizers and regularization. I would not recommend rejection, but the revision should address these issues before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is worth knowing: instead of pretraining a GNN transfer operator, they fit it per instance against the target-level loss. That is new relative to ODIL and learned-AMG lines. The matched LUNDIsim ablation gives credible evidence that the learned transfer beats simple interpolation by roughly 8–9% at the same fine-grid budget. That is the real result, and it holds up.\n\nWhat the paper does well: the framework is clean, keeps the prescribed numerical operator, needs no surrogates or offline data, and is demonstrated on two genuinely different inverse problems. No circularity in the main loop—the GNN is fitted to the target loss, not to ground-truth errors. The c_S calibration only sets a regularization weight, so that is minor.\n\nWhere it gets soft: the abstract's 1.25x number is misleading. Equation (38) counts only DNF solver steps. The 3000-step corrector fit at the 128^2 target grid adds about 0.5 to the normalized proxy, making the true ratio roughly 1.75x for the two-level Darcy run. For the four-level LUNDIsim run, with three transfers, the undercount is larger. The qualitative conclusion—large error reduction for moderate extra cost—likely survives the correction, but \"similar cost\" is not supported. The authors should re-report the cost with transfer fitting included.\n\nThe EIT comparison is weaker and confounded: the single-level DNF baseline uses Adam with 10,000 updates, while the complete pipeline uses L-BFGS with a different budget and a different regularization schedule. The mIoU gain over the official linearized CEM reconstruction is modest (0.603 to 0.623). This section reads as a breadth demonstration rather than a controlled claim, which the authors partly acknowledge. Also, they do not report unstructured-mesh experiments, and the paper notes that as a limitation.\n\nBottom line: this deserves a serious referee. The cost accounting needs correction, the EIT baseline should be matched, and code or full hyperparameter lists should be released. With those changes, the per-instance objective-fitted transfer is a solid algorithmic contribution.","headline":"Genuinely new per-instance learned transfer for multilevel inverse problems with a credible matched ablation—but the headline cost number omits corrector fitting, so the efficiency claim is overstated.","tokens_in":25032,"tokens_out":2055,"would_cite":true,"duration_ms":20266,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65N21","65N55","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"A learned coarse-to-fine transfer over nested discrete neural fields recovers PDE coefficients more accurately than optimizing directly on the finest grid, at modest extra grid work.","keywords":["PDE-constrained inverse problems","discrete neural fields","multilevel optimization","graph neural networks","learned prolongation","Darcy flow","electrical impedance tomography"],"falsifier":"Run the direct 128 Darcy solve with enough extra Adam steps to match the multilevel path's total Adam-step budget (about 10,500 steps at the fine-grid-equivalent cost, including the 3,000 transfer steps). If its error drops to the multilevel level, the claimed advantage is a compute-budget effect rather than a property of the coarse-to-fine path.","tokens_in":23995,"feed_emoji":"🧮","tokens_out":10436,"duration_ms":94416,"temperature":0.7,"pith_summary":"The paper claims that PDE-constrained inverse problems can be solved more accurately by optimizing progressively on nested grids with a learned, per-instance transfer between levels, rather than by optimizing directly on the finest grid. Each level is a discrete neural field solve: the discretized state and coefficient arrays are optimized directly under the fixed numerical model, with no surrogate or offline training. Between levels, a graph neural network (ResiPhy-GNN) starts from bilinear interpolation and learns small weight adjustments and a bounded residual correction by minimizing the target-level physics and data loss. In a controlled Darcy permeability experiment, the two-level 64-to-128 path uses 1.25x the cumulative grid-work proxy of the direct 128 solve while achieving 6.76x lower permeability error and 10.4x lower state error. On measured EIT data, the multilevel pipeline lifts the mean thresholded three-class overlap from 0.603 for the official linearized reconstruction to 0.623, and from 0.533 for direct single-level DNF.","feed_headline":"6.76x lower permeability error at 1.25x grid work","feed_subtitle":"Coarse-to-fine solver with per-instance learned transfer beats direct fine-grid inversion.","key_machinery":"The load-bearing component is ResiPhy-GNN, a graph neural network performing learned coarse-to-fine prolongation. Each fine node aggregates from a fixed coarse stencil using an interpolation-centered softmax: the prescribed bilinear weights are multiplied by learned edge-logit exponentials and renormalized, so the all-zero case recovers bilinear interpolation exactly. The aggregated value is then updated by a bounded correction (epsilon times tanh of a residual). A shared two-layer SiLU MLP, applied pointwise to node features (coordinates, interpolated baseline, observation and source channels, and an optional loss-gradient diagnostic), outputs the logit biases and the residual; the output l","core_discovery":"The central claim is that a per-instance, objective-fitted graph-network transfer (ResiPhy-GNN) over nested discrete neural field levels yields better reconstructions than direct single-level DNF optimization at similar cost. The transfer starts from bilinear interpolation, then learns edge-logit biases and a bounded nodewise correction by minimizing the target-level physics and data loss. In a controlled Darcy test, the 64-to-128 path uses 1.25x the cumulative grid-work proxy of the direct 128 solve while achieving 6.76x lower permeability error and 10.4x lower state error. On measured EIT data, the same construction improves the mean thresholded three-class overlap from 0.603 to 0.623. No","pith_inferences":["Including the 3000 Adam transfer-fitting steps (assessed at the target grid) in the cost accounting would raise the reported 1.25x work proxy to roughly 1.75x for one interface and more for three; the efficiency-precision trade-off likely remains favorable, but a full wall-clock comparison would settle it.","The per-instance fitting makes the method expensive when many levels are stacked; warm-starting or sharing corrector parameters across similar problems could cut that overhead, at the cost of the paper's deliberate no-pretraining design.","The residual update is bounded and the weights are interpolation-centered, suggesting the learned correction acts as a form of instance-adaptive regularization; one could test whether it degrades gracefully as observation noise increases, since the transfer loss includes the noisy data.","The graph-based formulation does not assume Cartesian structure, so the framework may extend to unstructured and adaptively refined meshes; the paper explicitly leaves this as future work, and a test on a triangular-mesh Darcy or EIT problem would be a natural check."],"forward_implications":["The error reductions persist over the full fine-grid optimization trajectory in the matched reservoir-like Darcy ablation, so the learned transfer improves the optimization path itself, not just the starting point.","The same multilevel construction works for both full-space Darcy inversion and reduced linearized EIT inversion, indicating the transfer module is not tied to a specific PDE, discretization, or observation geometry.","Because the transfer is fitted per instance using the prescribed numerical operator, the framework inherits the physics model directly and avoids the risk of surrogate errors being absorbed into the reconstructed coefficient.","The complete four-level run shows the approach scales to three interfaces while reducing the normalized PDE residual at each transition, with the largest reduction at the final coarse-to-fine step."],"fun_headline_variants":["6.76x lower permeability error at 1.25x grid work","Per-instance learned transfer beats direct fine-grid solving","No offline pretraining: multilevel neural field for inverse problems","Same multilevel scheme boosts Darcy and EIT inversions","Learn coarse-to-fine map on the fly, no surrogates"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The reported 1.25x cost factor counts only the DNF solver steps on each grid and omits the 3000 Adam steps spent fitting the GNN transfer at each interface; if the transfer cost is included, the true overhead is roughly 1.75x for the one-interface Darcy path (and larger for three-interface runs), so the strength of the 'similar cost' claim depends on this accounting.","fun_headline_variants_meta":{"raw":{"variants":["6.76x lower permeability error at 1.25x grid work","Per-instance learned transfer beats direct fine-grid solving","No offline pretraining: multilevel neural field for inverse problems","Same multilevel scheme boosts Darcy and EIT inversions","Learn coarse-to-fine map on the fly, no surrogates"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000729,"raw_usage":{"total_tokens":3162,"prompt_tokens":864,"completion_tokens":2298,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":608,"completion_tokens_details":{"reasoning_tokens":2210}},"tokens_in":608,"tokens_out":2298,"duration_ms":13930,"temperature":1.0,"reasoning_tokens":2210,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T08:16:19.728936+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the direct 128 Darcy solve with enough extra Adam steps to match the multilevel path's total Adam-step budget (about 10,500 steps at the fine-grid-equivalent cost, including the 3,000 transfer steps). If its error drops to the multilevel level, the claimed advantage is a compute-budget effect rather than a property of the coarse-to-fine path.","supporting_citations":[],"review_version":2}