{"id":"b955cf81-0d4e-4411-912b-46a8c5d316e3","arxiv_id":"2607.14665","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Training a single PINN on progressively finer collocation point sets, carrying over network parameters, cuts training time by an order of magnitude on tested Poisson, convection-diffusion-reaction, and Helmholtz benchmarks with comparable accuracy.","lead":"This paper proposes a training trick for physics-informed neural networks: start with few collocation points, then add more while reusing the already-learned network parameters. On several PDE benchmarks, it reports around 10-30x faster training with comparable accuracy.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No ablation isolates parameter transfer from the multilevel schedule, so the mechanism behind the reported speedups is unproven; a random-reinitialization control is needed.","rationale":"Reader's verdict is CONDITIONAL, and the stress-test agrees with the identified weakest assumption: the mechanism of coarse-to-fine parameter transfer is not isolated by any experiment. I reviewed the full text, including the incomplete Section 3.5 placeholder and the missing 3D example, but those are scope/completeness issues rather than the core argument's soft spot. The load-bearing concern is that the reported speedups are consistent with a simpler explanation: most optimization happens on very small collocation sets, and the trained network generalizes to the larger sets well enough that the final levels need almost no training (Table 3). A random-reinitialization ablation and a coarse-only baseline would settle whether the warm start and the refinement hierarchy are actually responsible. If they are not, the central claim about the strategy's mechanism is unsupported, though the raw empirical speedups might still hold. This does not change the CONDITIONAL assessment; it sharpens the condition under which the paper should be accepted.","tokens_in":11998,"tokens_out":11022,"duration_ms":110153,"concrete_test":"On Example 4.1 (and, if possible, Example 5.2), implement three controls with identical point sets, tolerances, optimizer, and architecture: (i) MPU-PINN exactly as in Algorithm 1; (ii) the same level schedule but re-initialize θ randomly (Glorot) at the start of every level before training; (iii) train only on the initial 81-point set and evaluate on the full grid without refinement. Report wall-clock time and relative L2 error for 5 seeds (mean±std). If (ii) is within ~20% of (i) in time, the speedup is not due to parameter transfer; if (iii) achieves relative error within a factor of 1.3 of the conventional PINN while being faster than (i), the hierarchical refinement itself is not load-bearing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central efficiency claim—that MPU-PINNs reduce training time by 10–30x (Tables 2, 4, 7, 11) because coarse-level parameters provide a favorable initialization for finer levels—is not supported by the experiments as reported. The paper contains no ablation that isolates parameter transfer from the multilevel schedule. In the nested scheme (Algorithm 1), the same tolerance ε is applied at every level, and Table 3 shows that for Example 4.1 levels 3–7 each terminate after 1 epoch. This means the coarse-level network already satisfies ε on the expanded point sets; the speedup is then dominated by the fact that most epochs are spent on 81–289 points, not by warm-starting per se. A control that re-initializes the network at each level and trains on exactly the same nested sets to the same ε would distinguish these. Without it, Section 3.4's cost model only proves that coarse epochs are cheaper; it does not prove that transfer reduces the number of fine-level iterations. The comparison protocol also differs across methods (e.g., 100k-epoch caps for baselines in Example 5.1 while MPU runs 132k total epochs), and no error bars are reported, so the magnitude of the claimed speedup is not robustly quantified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces MPU-PINNs, a training strategy for physics-informed neural networks that starts with a coarse set of collocation points and progressively refines it, transferring the trained parameters to each finer level. A loss-scaling reformulation is also proposed for high-frequency problems. The authors benchmark against conventional PINNs, SA-PINNs, and XPINNs on 2D Poisson, convection–diffusion–reaction, and Helmholtz problems, reporting roughly 10–30x reductions in wall-clock time with comparable accuracy, and claim extension to 3D Poisson.","tokens_in":12316,"tokens_out":5002,"duration_ms":47212,"significance":"If the reported speedups generalize, MPU-PINNs are a simple, appealing drop-in acceleration for single-network PINN training. The paper provides explicit algorithms (Algorithms 1 and 2), a cost model in Section 3.4, and consistent empirical evidence across four 2D benchmarks, including epoch distributions that make the coarse-to-fine behavior transparent. The scaling strategy for high-frequency problems is a useful contribution. However, the manuscript currently contains unfinished text and overclaims 3D results, and the mechanism of the speedup is not isolated experimentally.","major_comments":[{"comment":"Section 3.5 is an unfinished placeholder. It contains the text 'EXPERIMENTS....', bullet points with 'Use examples (test problems) in PIP-IO', and the speculative sentence 'MPU probably has an advantage...' with no results. This is not a completed section and should be either fully written or removed. Its presence makes the manuscript unsuitable for publication in its current form.","section":"Section 3.5"},{"comment":"The abstract and conclusion state that MPU-PINNs are evaluated on 'two- and three-dimensional Poisson equations', but all numerical examples (4.1, 4.2, 5.1, 5.2) are set in Ω=(0,1)^2. No 3D experiment appears in the paper. The claimed scope overstates the evidence; the authors should either add a 3D Poisson experiment or revise the claims to match the actual experiments.","section":"Abstract and Section 6"},{"comment":"The central explanatory claim—that parameter transfer from coarse to fine levels is what reduces fine-level iterations—is not isolated. Section 3.4's cost model only shows that coarse epochs are cheaper (O(N_l P)); it does not show that transfer reduces the number of fine-level iterations. The observed decreasing epochs per level could also follow from applying the same stopping tolerance ε to nested point sets, with several fine levels terminating after 1 epoch. A control experiment that re-initializes the network at each level (or otherwise ablates warm-starting) is necessary to support the claimed mechanism. Without this, the speedup could be attributed to the multilevel schedule or the stopping criterion rather than to the transferred parameters.","section":"Section 3.4 and Tables 3, 5, 6, 9, 10"}],"minor_comments":[{"comment":"'The time required ... is nearly 5% shorter' is inconsistent with Table 2, where MPU time is about 3% of the conventional PINN time (51.5 s vs 1650.98 s). Should read 'less than 5% of the elapsed time'.","section":"Section 4.1, text near Table 2"},{"comment":"'PIP-IO' is undefined and appears to be a leftover note.","section":"Section 3.5"},{"comment":"Typos and wording issues: 'leff' (Section 5), 'convectio–diffusion–reaction' (Section 4 intro), 'The another settings' (Section 5.1), 'Figure 2 exhibits' etc. A careful proofread is needed.","section":"Throughout"},{"comment":"The symbol L is used both for the level counter and the loss value; this is confusing. Rename the loss, e.g., Loss.","section":"Algorithms 1 and 2"},{"comment":"Baselines are capped at 100,000 epochs while MPU-PINNs are reported with total epochs exceeding 130,000 (e.g., Algorithm 1: 99099+33372.8+1+85). The paper should clarify the per-method stopping criteria and whether the cap applies identically.","section":"Example 5.1, Table 6"},{"comment":"No standard deviations or confidence intervals are reported despite five independent runs; adding these would strengthen the robustness of the accuracy and timing comparisons.","section":"Section 4, experimental protocol"}],"recommendation":"major_revision","confidential_remarks":"This manuscript appears to be an early draft: Section 3.5 contains placeholder text and the abstract claims 3D results that are absent. The editor may wish to verify that this is the intended submission and whether the authors plan to complete the manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The practical upshot is real: on the 2D benchmarks they actually ran, MPU-PINNs cut wall-clock time by roughly 10–30x while staying within a factor of about 1.3 of the baselines' relative errors. That's a worthwhile empirical result for the PINN community, and the coarse-to-fine strategy is a natural extension of multigrid/curriculum intuitions, clearly spelled out in Algorithms 1 and 2.\n\nThe problems are with execution and evidence. Section 3.5 is literally an unfinished placeholder ('EXPERIMENTS....' and bullets). The abstract and conclusion promise three-dimensional Poisson results, but the experiments stop at 2D. There is no code and no error bars, though the method is described well enough to reimplement.\n\nThe larger scientific gap is the missing ablation. Nothing isolates parameter transfer from the multilevel schedule itself. Table 3 shows levels 3–7 each finishing in 1 epoch in Example 4.1, which means the model trained on 289 points already satisfies the tolerance on the 1,089-point set. That's consistent with the speedup being mostly about cheap coarse epochs, not warm-starting. A control that randomly reinitializes the network at each level, trained on the same nested sets to the same tolerance, would settle the question. Until that runs, the claimed mechanism—'parameters obtained at a coarse level provide a favorable initialization'—is just an assumption.\n\nAlso, the stopping tolerance and initial level are hand-picked per problem and are part of the speed comparison. The reference list omits most of the multigrid/curriculum/transfer learning literature, which is a minor blemish.\n\nMy verdict: this is a promising preprint, not a publishable paper. The core efficiency claim for 2D is plausible and worth pursuing, but the placeholder and the missing 3D results make it hard to take seriously as a submission. I'd send it to reviewers only if you want them to name exactly those missing pieces; otherwise, desk reject and invite resubmission with the ablation included.","headline":"Useful 2D results, but the paper is unfinished and the transfer mechanism is unproven.","tokens_in":12796,"tokens_out":4936,"would_cite":false,"duration_ms":51877,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65N35","65N50","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"Training a physics-informed neural network first on few collocation points, then transferring the learned parameters as the point set grows, cuts training time by an order of magnitude or more on several benchmark PDEs.","keywords":["physics-informed neural networks","multigrid training","parameter transfer","coarse-to-fine learning","collocation points","Helmholtz equation","spectral bias","convection-diffusion-reaction"],"falsifier":"Run a single-network PINN on a problem where the coarse grid cannot represent the solution's structure (e.g., a sharp interior layer or a high wavenumber with an initial level too coarse), and measure epochs at the finest level. If the final level still requires as many iterations as a from-scratch run — or if the loss decreases at the same rate regardless of whether the initialization comes from a coarser level — then coarse-to-fine parameter transfer is not the source of the speedup, and the reported gains would instead be attributable to the stopping tolerance or epoch cap.","tokens_in":11887,"feed_emoji":"⚡","tokens_out":8974,"duration_ms":90596,"temperature":0.7,"pith_summary":"MPU-PINNs is a training strategy for physics-informed neural networks that starts with a small set of collocation points, trains the network to a residual tolerance, then transfers the trained parameters to a denser set, repeating until the target resolution is reached. The central claim is that coarse-to-fine initialization lets the network capture global solution structure cheaply, so the expensive finest-level optimization needs far fewer epochs than training from random initialization. Across the benchmarks reported — 2D Poisson, convection-diffusion-reaction, high-frequency Poisson, and Helmholtz — this yields roughly 10–30x shorter wall-clock training times than conventional PINNs while keeping relative errors comparable, with a scaled loss formulation added for oscillatory cases. The method keeps a single network throughout, so the speedup comes from the training schedule and parameter transfer, not from model size or domain decomposition. If the mechanism holds, any single-network PINN implementation can be accelerated by changing only how collocation points are introduced.","feed_headline":"Coarse-to-fine training cuts PINN training time up to 30x","feed_subtitle":"Start with few collocation points, transfer parameters as points grow: 10-30x faster with comparable error.","key_machinery":"The central object is the multilevel collocation-point schedule with parameter transfer: a sequence of point sets N_1 < N_2 < ... < N_L ending at the target resolution, with the optimized parameters at level ℓ used to initialize level ℓ+1. The cost identity is Cost_MPU ≈ P Σ_ℓ K_ℓ N_ℓ versus Cost_PINN ≈ K_f N_f P, so savings are realized whenever coarse levels are cheap and the finest level needs far fewer iterations. Two point-selection variants are offered: nested addition, where new points are added to the existing set, and random resampling. The supporting mechanism for oscillatory problems is a rescaled loss that divides the Laplacian term by k^2, making the residual scale comparable to","core_discovery":"On the paper's own terms, the discovery is that most of the optimization work in PINN training can be moved from the finest, most expensive collocation set to coarse sets without sacrificing accuracy. Trained on a coarse set first, the network parameters already approximate the low-frequency global structure of the solution; when those parameters initialize the next level, the method reaches the fine-level stopping tolerance in a handful of epochs, whereas conventional PINNs require tens of thousands. The concrete demonstration is in the epoch tables and elapsed-time comparisons: the finest level consumes 1–85 epochs for MPU-PINNs versus 8,344–100,000 for the baselines. For high-frequency pr","pith_inferences":["A straightforward extension the paper leaves untested is applying the same schedule to time-dependent and inverse PDEs; the paper's cost model predicts comparable or larger savings there because residual evaluation cost grows with the number of collocation points rather than with model complexity.","An ablation that fixes the final loss threshold rather than a per-level tolerance and equalizes total compute would isolate whether the speedup comes from coarse-level initialization quality or from early stopping built into the level scheduler.","The method could be combined with residual-based adaptive point selection, adding points where the residual is largest at each level, which the paper does not explore.","The abstract promises a 3D Poisson experiment and Section 3.5 announces a comparison with residual-weighted PINNs, but neither appears in the body text; the current numerical evidence is therefore limited to the reported 2D cases."],"forward_implications":["Across the reported 2D Poisson, convection-diffusion-reaction, high-frequency Poisson, and Helmholtz benchmarks, MPU-PINNs cut elapsed training time to roughly 3–10% of conventional PINN time while keeping relative errors comparable.","The finest-level epoch counts fall from thousands or tens of thousands (8,344 for 2D Poisson, 31,848 for convection-diffusion-reaction, 97,730 for Helmholtz) to a handful, often 1–85 epochs.","Both nested and random point-selection variants produce the speedup; nested selection tends to be faster, random selection slightly more accurate on the 2D Poisson case.","The scaled loss formulation (−(1/k^2)Δu = f̃) reduces relative error from about 2.3e-1 to 2.2e-2 in the high-frequency Poisson test and further cuts training time, showing that the PDE reformulation matters for oscillatory solutions.","The framework uses a single network with no subdomain interface conditions, so the acceleration is attributed to the training schedule rather than to model size or domain decomposition."],"fun_headline_variants":["Multigrid-based PINNs: train on coarse points first, then refine","Coarse-to-fine collocation: a faster way to train physics-informed nets","Hierarchical collocation refinement speeds up PINN training","Start with a few collocation points, then scale up: faster PINN training","Coarse-to-fine point refinement cuts PINN training cost"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The whole efficiency argument rests on the Section 3 assertion that parameters trained on a coarse collocation set 'typically provide a favorable initialization' for the next finer level; this premise is not proved, and the text's own missing pieces — no 3D experiment despite the abstract and an empty Section 3.5 with an announced comparison — mean the empirical support is currently limited to the 2D benchmarks actually reported.","fun_headline_variants_meta":{"raw":{"variants":["Multigrid-based PINNs: train on coarse points first, then refine","Coarse-to-fine collocation: a faster way to train physics-informed nets","Hierarchical collocation refinement speeds up PINN training","Start with a few collocation points, then scale up: faster PINN training","Coarse-to-fine point refinement cuts PINN training cost"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001535,"raw_usage":{"total_tokens":6010,"prompt_tokens":803,"completion_tokens":5207,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":547,"completion_tokens_details":{"reasoning_tokens":5112}},"tokens_in":547,"tokens_out":5207,"duration_ms":36373,"temperature":1.0,"reasoning_tokens":5112,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T01:25:52.277872+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a single-network PINN on a problem where the coarse grid cannot represent the solution's structure (e.g., a sharp interior layer or a high wavenumber with an initial level too coarse), and measure epochs at the finest level. If the final level still requires as many iterations as a from-scratch run — or if the loss decreases at the same rate regardless of whether the initialization comes from a coarser level — then coarse-to-fine parameter transfer is not the source of the speedup, and the reported gains would instead be attributable to the stopping tolerance or epoch cap.","supporting_citations":[],"review_version":1}