{"id":"3d0fdbe0-3b4a-4245-a8cb-dd236adb9e37","arxiv_id":"2607.06091","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":8,"one_line_summary":"Adding a soft mass-conservation penalty to PINNs for the 1D advection-diffusion equation reduces long-term relative L2 error by 9–67× and mass error by 15–215× compared to vanilla PINNs across Peclet numbers 0.01–20.","lead":"This paper adds a soft mass-conservation penalty to physics-informed neural networks (PINNs) for the 1D advection-diffusion equation and shows it reduces long-term mass drift and accuracy loss compared to standard PINNs. A smart generalist might read it to understand a practical, low-overhead fix for conservation-law violations in neural PDE solvers.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"Headline improvement factors (9–67×, 15–215×) are reported as point estimates from apparently single training runs with no variance across random seeds, leaving the quantitative claims statistically unsubstantiated.","rationale":"The reader's verdict of CONDITIONAL with MODERATE confidence is appropriate. The paper demonstrates a legitimate and useful contribution — systematically evaluating a soft mass-penalty constraint for PINNs on the advection-diffusion equation across Peclet numbers — and the qualitative finding (mass penalty helps long-term accuracy by reducing mass drift) is plausible and well-motivated by the NTK/gradient pathology literature cited (§4). However, the quantitative claims are under-evidenced: no error bars, no seed variation, and an underspecified tuning procedure. The reader correctly identified several of these gaps (error bars, λ_m sensitivity) in their rationale, though their formal weakest_assumption over-weighted the analytical-mass-target concern, which is not actually a limitation for forward problems with known initial conditions. The reader's mention of broken citations (§2.4 has '[?]' for the activation function citation, §1 has '[?]' for integral PINN) is also valid but minor. The verdict should remain CONDITIONAL: the technique is sound and the contribution is real, but the quantitative claims need statistical backing before they can be fully accepted. No adjustment to the verdict is needed — the reader already captured the right level of caution.","tokens_in":8666,"tokens_out":2683,"duration_ms":101387,"concrete_test":"Re-run all 7 Peclet number configurations for both Vanilla and Mass-Penalty PINN with at least 5 independent random seeds (varying network initialization and LHS collocation sets). Report mean ± std for both relative L2 error and mass conservation error at T=100s. If the coefficient of variation for the improvement factor exceeds ~30% at any Peclet number, the headline ranges (9–67×, 15–215×) are not robust. Additionally, clarify whether §2.5 hyperparameter tuning was performed for Vanilla or Mass-Penalty configuration; if for Mass-Penalty, re-tune Vanilla independently and recompute the comparison.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on specific quantitative improvement factors, but PINN training is stochastic: network initialization, collocation point sampling (via LHS), and optimizer trajectories all vary across seeds. The paper reports no error bars, no standard deviations, and no mention of how many random seeds were used (§2.5, §3.2). The improvement ranges (9–67× for L2, 15–215× for mass error) are presented as deterministic facts. If even one Peclet number's improvement factor has high variance across seeds — say, Pe=0.01 where the Mass-Penalty PINN slightly underperforms Vanilla in short-term runs (§3.1) — the lower bound of the claimed improvement range could be unreliable or non-reproducible. This is compounded by an underspecified hyperparameter tuning procedure (§2.5): the penalty weights (λ_IC=100, λ_BC=20, λ_PDE=10, λ_m=10) were established via a 'preliminary' study [29], but it is unclear whether this optimization was performed with the mass penalty active or for the Vanilla configuration. If weights were tuned with the mass penalty on, the Vanilla PINN baseline may be using suboptimal hyperparameters, inflating the apparent improvement. The reader's weakest_assumption focused on the analytical mass target and λ_m sensitivity; the analytical mass concern is less load-bearing than it appears, since the initial mass is computable from the given initial condition (Eq. 3) for any forward problem. The λ_m sensitivity concern is valid but secondary to the more fundamental issue of no statistical validation.","agreement_with_reader":"partial"},"referee_report":{"model":"glm-5.2","summary":"This manuscript evaluates a Mass-Penalty PINN for the 1D advection-diffusion equation, comparing it against a Vanilla PINN and a Crank-Nicolson baseline across seven Peclet numbers (0.01–20) and two time horizons (T=5 s, T=100 s). The mass-penalty approach augments the standard PINN loss with a soft constraint penalizing deviation from the analytically known initial mass. The central claim is that for long-term simulations (T=100 s), the Mass-Penalty PINN reduces relative L2 error by 9–67× and mass conservation error by 15–215× compared to Vanilla PINN, at a training cost increase of only 5.4%. The experimental design is reasonable: the Peclet range spans diffusion- to advection-dominated regimes, an analytical reference solution is available, and a grid-independence study establishes the CN baseline.","tokens_in":8952,"tokens_out":1315,"duration_ms":119459,"significance":"The paper addresses a practically relevant problem—mass drift in PINNs over long integration horizons—and provides a systematic sweep over Peclet numbers that is uncommon in the PINN literature. The use of an analytical reference solution and a grid-verified CN baseline is a strength. The mass-penalty approach itself is not novel (it follows [28]), but its systematic evaluation for advection-diffusion transport across flow regimes fills a documented gap. The computational cost analysis (Table 1) is a useful addition. However, the quantitative claims rest on experimental results whose statistical robustness is not established, which limits the significance of the reported improvement factors.","major_comments":[{"comment":"§2.5, §3.2: All quantitative results (improvement factors of 9–67× for L2 error and 15–215× for mass error) appear to be from single training runs. No error bars, standard deviations, or number of random seeds are reported anywhere in the manuscript. PINN training is stochastic: network initialization, LHS collocation point sampling, and optimizer trajectories all vary across seeds. The headline improvement factors are presented as deterministic ranges, but without variance estimates it is impossible to assess whether the lower bounds (9×, 15×) are reproducible or whether they could collapse for specific Peclet numbers. This is load-bearing for the central claim. At minimum, the authors should report results averaged over multiple seeds (≥3–5) with standard deviations for each Peclet number and time horizon, and revise the improvement-factor claims accordingly.","section":null},{"comment":"§2.5: The penalty weights (λ_IC=100, λ_BC=20, λ_PDE=10, λ_m=10) were established via a 'preliminary hyperparameter tuning study' citing [29] (DeepXDE). It is unclear whether this tuning was performed with the mass penalty active or for the Vanilla configuration only. If the weights were optimized with the mass penalty on, the Vanilla PINN baseline may be using suboptimal hyperparameters, which would inflate the apparent improvement. The authors should clarify which configuration was used during tuning, and ideally confirm that the Vanilla baseline uses its own optimized weights. Additionally, the sensitivity of results to λ_m=10 is not reported; a brief sensitivity analysis (e.g., λ_m ∈ {1, 10, 100}) would strengthen the claim that a single fixed λ_m is robust across all Peclet numbers.","section":null}],"minor_comments":[{"comment":"§2.4, Eq. (8): The notation switches between L_m and L_mass (§3.1 uses 'Lmass constraint with λmass = 10'). Standardize on one notation.","section":null},{"comment":"§2.4, Eq. (9): The Gaussian quadrature rule used for the mass integral is not specified (number of points, order). Since this integral is evaluated at every training step and its accuracy directly affects the mass-penalty term, the quadrature details should be stated.","section":null},{"comment":"§2.4: Reference [26] (and possibly [?]) appears with a question mark in the reference list, indicating a missing citation. Please fix.","section":null},{"comment":"§3.1: The statement that the Mass-Penalty PINN 'slightly underperforms' Vanilla in diffusion-dominated short-term runs is qualitative. Quantifying the difference (e.g., relative L2 error values for Pe=0.01 at T=5 s for both methods) would help the reader assess the tradeoff.","section":null},{"comment":"Figures 3–7: The axis labels and legends are small and difficult to read. Consider enlarging font sizes and ensuring line styles/colors are distinguishable in print.","section":null},{"comment":"§3.3, Table 1: The CN inference time (≈3.03 s) is not directly comparable to PINN inference time because CN solves the full spatio-temporal grid while the PINN evaluates at queried points. A brief note clarifying what 'inference' means for each method would improve fairness of the comparison.","section":null},{"comment":"§5: The conclusion states the mass penalty reduces errors 'by up to 67 and 215 times,' but the abstract and §3.2 report ranges (9–67, 15–215). Ensure consistency in phrasing.","section":null}],"recommendation":"major_revision","confidential_remarks":"The core idea is sound and the experimental design is reasonable for a computational physics paper. The main concern is the absence of any statistical validation of the headline improvement factors, which is standard practice when reporting PINN benchmarks. If the authors can provide multi-seed results and clarify the hyperparameter tuning procedure, the paper should be publishable. The analytical-mass-target concern raised in the reader's report is less concerning than it appears: for the forward problem studied here, the initial mass is computable from the given initial condition (Eq. 3), so the target is not circular. The applicability to inverse problems without known analytical solutions is a legitimate scope limitation but does not undermine the forward-problem claims."},"author_rebuttal":{"model":"glm-5.2","summary":"We thank the referee for a careful and constructive report. Both major comments are well-taken and address genuine gaps in the experimental rigor of the manuscript. We address each below.","responses":[{"response":"The referee is correct. All results in the current manuscript are from single training runs, and no variance estimates are reported. This is a genuine weakness in the experimental design, and we agree that the improvement factors (9–67× for L2 error, 15–215× for mass error) cannot be reliably interpreted without knowing the spread across random seeds. We will revise the manuscript to include results averaged over at least 5 random seeds for every Peclet number and both time horizons, reporting mean ± standard deviation for both the relative L2 error and the mass conservation error. We will also recompute and revise the improvement-factor ranges accordingly, reporting them as mean improvement factors with appropriate uncertainty rather than as deterministic bounds. If any Peclet number shows high variance that undermines the lower-bound claims, we will state this explicitly rather than reporting only the favorable cases. The abstract and conclusions will be updated to reflect the revised, statistically grounded claims.","revision_made":"yes","referee_comment":"§2.5, §3.2: All quantitative results appear to be from single training runs. No error bars, standard deviations, or number of random seeds are reported. PINN training is stochastic. The headline improvement factors are presented as deterministic ranges, but without variance estimates it is impossible to assess whether the lower bounds (9×, 15×) are reproducible. At minimum, report results averaged over multiple seeds (≥3–5) with standard deviations for each Peclet number and time horizon, and revise the improvement-factor claims accordingly."},{"response":"We acknowledge that the manuscript is unclear on this point. In the current version, the penalty weights (λ_IC=100, λ_BC=20, λ_PDE=10) were tuned on the Vanilla PINN configuration first, and then λ_m=10 was selected separately for the Mass-Penalty PINN. However, this procedure was not documented in the manuscript, and the referee is right that the lack of clarity raises the concern that the Vanilla baseline may be disadvantaged (or, conversely, that the Mass-Penalty configuration was not jointly optimized). We will revise §2.5 to explicitly state the tuning protocol. In the revised manuscript, we will also conduct and report a separate hyperparameter tuning for the Vanilla PINN to confirm that its weights are not suboptimal. Regarding the sensitivity to λ_m: the referee's suggestion is reasonable and we agree it strengthens the robustness claim. We will add a sensitivity analysis for λ_m ∈ {1, 10, 100} across representative Peclet numbers (at minimum Pe = 0.01, 1, 20) for the long-term horizon, reporting both L2 and mass errors. If the results show that λ_m=10 is not robust across all regimes, we will report this honestly and discuss regime-dependent choices.","revision_made":"yes","referee_comment":"§2.5: The penalty weights were established via a 'preliminary hyperparameter tuning study' citing [29] (DeepXDE). It is unclear whether this tuning was performed with the mass penalty active or for the Vanilla configuration only. If the weights were optimized with the mass penalty on, the Vanilla PINN baseline may be using suboptimal hyperparameters, which would inflate the apparent improvement. Clarify which configuration was used during tuning, and ideally confirm that the Vanilla baseline uses its own optimized weights. Additionally, the sensitivity of results to λ_m=10 is not reported; a brief sensitivity analysis (e.g., λ_m ∈ {1, 10, 100}) would strengthen the claim that a single fixed λ_m is robust across all Peclet numbers."}],"tokens_in":8388,"tokens_out":1048,"duration_ms":40044,"standing_objections":[]},"desk_editor":{"model":"glm-5.2","letter":"The paper applies a soft mass-conservation penalty (from Huang et al. [28], originally for Cahn-Hilliard) to the 1D periodic advection-diffusion equation and systematically evaluates it across seven Peclet numbers and two time horizons. The headline finding is that vanilla PINNs accumulate significant mass drift over long integration times (T=100s, 14–21% mass loss), and adding the penalty term reduces both L2 error and mass error by large factors while costing only 5.4% more training time. The technique itself is not new, but its systematic evaluation on advection-diffusion across Peclet regimes is a legitimate contribution. The paper is honest about the short-term results, where the mass penalty sometimes slightly hurts accuracy in diffusion-dominated regimes — that candor is welcome. The analytical reference solution, grid-independence study for the CN baseline, and computational cost analysis are all done properly. The core qualitative claim — that mass drift is the dominant failure mode for vanilla PINNs in long-time conservative transport, and that a soft penalty mitigates it — is well-supported by the evidence shown. The correlation between L2 error and mass error in Figure 7 is a nice diagnostic. The discussion of spectral bias and Kolmogorov n-width limitations in advection-dominated regimes is appropriate. The biggest soft spot is the absence of error bars or any mention of multiple random seeds. PINN training is stochastic — initialization, LHS sampling, optimizer trajectory all vary — and the improvement factors (9–67×, 15–215×) are presented as point estimates from apparently single runs. This doesn't undermine the qualitative finding, but it makes the specific numbers unreliable. A referee should insist on at least 3–5 seeds with reported variance. Second, the hyperparameter tuning procedure (§2.5) is underspecified: it's unclear whether the penalty weights were tuned with the mass term active, which could inflate the vanilla baseline's disadvantage. A λ_m sensitivity sweep would also help. The stress-test note raised the analytical-mass-target concern, but I think that's less serious than it sounds — the initial mass is computable from the given initial condition for any forward problem, so it's not circular. The two broken citations ([?] in §1 and §2.4) are minor but should be fixed. This paper is for PINN practitioners working on conservative transport problems. It won't change the field, but it's a careful, honest benchmark study. It deserves a serious referee who can push for statistical validation and hyperparameter transparency.","headline":"Solid empirical study applying a known mass-penalty technique to 1D advection-diffusion; main gap is no variance reporting across seeds.","tokens_in":9477,"tokens_out":1064,"would_cite":false,"duration_ms":29299,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["02.60.Cb","02.70.-c","47.11.-j"],"model":"glm-5.2","headline":"Neural Net Solver Gets Mass Conservation Right at 5% Cost","keywords":[],"falsifier":"If a PINN trained with the mass-penalty constraint on a long-horizon advection-diffusion problem still exhibited mass drift comparable to the Vanilla PINN (e.g., >10% mass loss), the central claim would be falsified.","tokens_in":8741,"feed_emoji":"⚖️","tokens_out":690,"duration_ms":47696,"temperature":0.7,"pith_summary":"This paper identifies accumulated mass drift as the primary failure mode of standard physics-informed neural networks (PINNs) when simulating the 1D advection-diffusion equation over long time horizons. The authors demonstrate that a standard PINN, which only penalizes local PDE residuals, can lose up to 21% of total mass over a 100-second simulation. To fix this, they evaluate a Mass-Penalty PINN that adds a soft global constraint to the loss function, anchoring the total integrated mass to its initial analytical value. The central finding is that this simple addition reduces relative L2 error by factors of 9 to 67 and mass conservation error by factors of 15 to 215 across all tested Peclet numbers, while increasing training time by only 5.4%.","feed_headline":"Adding Mass Conservation to AI Solvers Cuts Error 67x at 5% Cost","feed_subtitle":"Standard neural PDE solvers lose up to 21% of mass over long simulations. A soft penalty constraint fixes this with negligible overhead.","key_machinery":"Mass-Penalty PINN: A physics-informed neural network augmented with a loss term that penalizes the squared difference between the analytically computed initial mass and the network's predicted spatial integral of mass at each time step, computed via Gaussian quadrature.","core_discovery":"The long-term accuracy degradation in Vanilla PINNs for conservative transport is shown to be predominantly caused by the accumulation of mass drift, not by an inability to learn the spatial profile. By introducing a soft penalty term that forces the network to preserve the global integral of the concentration field, the Mass-Penalty PINN effectively eliminates this drift, restricting mass deviation to under 0.5% and stabilizing the L2 error on the order of 10^-3 over extended integration horizons.","pith_inferences":[],"forward_implications":["Mass-penalty constraints could be applied to other conservative PDEs (e.g., Navier-Stokes, Maxwell's equations) to improve long-term stability of neural network solvers.","The finding that local PDE residuals are insufficient for global conservation suggests a broader redesign of PINN loss functions to include integral invariants by default.","For inverse problems where analytical mass is unknown, empirically estimated mass from data could potentially substitute the analytical target, though this remains untested in the paper.","The 5.4% training overhead suggests that adding conservation constraints is computationally cheap relative to the accuracy gains, making it a practical add-on for existing PINN implementations."],"fun_headline_variants":["Soft mass-conservation constraint stops long-term PINN drift","A soft mass penalty eliminates long-term PINN accuracy loss","Mass-conserving penalty cuts long-term PINN error by 67x","Fixing mass drift improves long-term PINN transport reliability","Global mass penalty stabilizes long-term advection-diffusion PINNs"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The mass-penalty term relies on knowing the exact initial total mass from an analytical solution, which would be unavailable in inverse problems or real-world scenarios without a known ground truth.","fun_headline_variants_meta":{"raw":{"variants":["Soft mass-conservation constraint stops long-term PINN drift","A soft mass penalty eliminates long-term PINN accuracy loss","Mass-conserving penalty cuts long-term PINN error by 67x","Fixing mass drift improves long-term PINN transport reliability","Global mass penalty stabilizes long-term advection-diffusion PINNs","Soft mass constraint restricts PINN mass drift below 0.5%"]},"model":"glm-5.2","effort":"high","cost_usd":0.0,"raw_usage":{"total_tokens":946,"prompt_tokens":594,"completion_tokens":352,"prompt_tokens_details":null},"tokens_in":594,"tokens_out":352,"duration_ms":9668,"temperature":1.0,"reasoning_tokens":259,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-08T16:45:20.046096+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"If a PINN trained with the mass-penalty constraint on a long-horizon advection-diffusion problem still exhibited mass drift comparable to the Vanilla PINN (e.g., >10% mass loss), the central claim would be falsified.","supporting_citations":[],"review_version":1}