{"id":"f8a2cefc-94da-4893-a55c-9ea65968b160","arxiv_id":"2502.07325","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A curriculum-then-transfer PINN training scheme extends accurate time-domain simulation from short windows to long horizons across three test problems.","lead":"This paper combines curriculum learning and transfer learning to make physics-informed neural networks solve time-dependent PDEs over long horizons by splitting the time axis into short intervals. It tests the idea on nonlinear wave propagation, a vibrating plate, and a real reservoir flow model, reporting better accuracy than a standard PINN.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The method's distinguishing mechanism is self-generated pseudo-labels, yet the paper provides no control or error bound for those labels, so the claimed long-horizon advantage rests on an unverified error-accumulation assumption.","rationale":"The reader's weakest assumption is the reliability of self-generated supervised points, and my reading converges on the same point. The paper's central claim is an empirical one: CTL-PINN achieves accurate and stable long-time simulations. The only mechanism that separates CTL-PINN from a standard time-marching or growing-domain PINN is the re-use of the previous network's predictions as ground truth, both as pseudo-labels and as initial conditions. That mechanism is exactly where the argument is least secure, because there is no independent error source to correct drift. The authors' own Conclusion (2) explicitly states that error accumulation occurs, which is an in-scope limitation statement and supports the concern. I considered other potential objections, such as the absence of multiple seeds or error bars, the lack of released code, and the heuristic step-size recommendations, but these are secondary and would not by themselves invalidate the central claim; they would only reduce confidence. The pseudo-label error-accumulation concern is more fundamental because it targets the load-bearing premise of the training schedule. A clean ablation replacing pseudo-labels with exact labels, plus a per-step record of pseudo-label error against the analytical solution in the overlap, would settle whether the reported final error is dominated by the self-generated labels. If the exact-label ablation is much more accurate, the method's advantage is contingent on pseudo-label accuracy; if not, the concern is refuted and the conditional verdict can be upgraded. Since the reader already made the pseudo-label issue the weakest assumption and issued a CONDITIONAL verdict, my independent stress test does not move the verdict; the condition should explicitly include the overlap-error ablation.","tokens_in":21347,"tokens_out":4670,"duration_ms":43021,"concrete_test":"Re-run Section 4.1 forward problem (T_p=20, T_c=50, ΔT_pc=5, ΔT_ct=10) and record, at every curriculum and transfer step, the L2 error on the overlap subdomain between the previous model's prediction (the pseudo-label) and the analytical solution; plot this against the final CTL-PINN error. Additionally, in one ablation, replace the pseudo-labels in Eqs. (13)/(18) and the transfer initial condition in Eq. (19) with exact analytical values and rerun. If the exact-label ablation gives materially lower final error, error accumulation from self-generated labels is the limiting factor and the central claim is only as strong as the pseudo-labels. If the errors are comparable, the concern does not land.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Section 5) is that CTL-PINN yields accurate, stable long-time solutions and outperforms standard PINN. The mechanism that distinguishes the method is the repeated use of the previous network's own predictions as 'additional supervised learning points' (Eqs. (13) and (18), losses in Eqs. (16) and (22)) and, in transfer learning, as the initial condition for the next window (Eq. (19)). These labels are not independent data: if the model at step k−1 has error e_{k−1}, the supervised loss at step k asks the new model to match a target carrying that error, and the transfer initial condition injects it directly. The paper itself acknowledges this in Conclusion (2): 'there is a certain error between it and the exact values, leading to some error accumulation.' The load-bearing premise is therefore that e_{k−1} stays small enough that the pseudo-label bias does not dominate the residual and boundary losses. No error bound, contraction estimate, or accumulation analysis is provided, and the reported final L2 errors do not distinguish error already present in the pseudo-labels from error newly introduced by the training step. If the premise fails, the claimed advantage over standard PINN is not a property of the method but of favorable parameter choices in three examples.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CTL-PINN, a curriculum-transfer-learning variant of physics-informed neural networks for long-time simulation of time-dependent PDEs. The method decomposes the time domain into short windows, trains a standard PINN on the first window, then repeatedly extends the domain by using the previous network's predictions as additional supervised data (Eqs. (13), (18)) and as initial conditions for the next transfer window (Eq. (19)). The authors claim that this combination overcomes the instability and poor accuracy of standard PINNs over long horizons, requires fewer steps than curriculum-only or transfer-only variants, and extends to inverse problems. The method is demonstrated on nonlinear wave propagation, Kirchhoff plate dynamics, and a Saint-Venant hydrodynamic model of the Three Gorges Reservoir Area, including roughness-coefficient inversion and 24-hour-ahead prediction.","tokens_in":21644,"tokens_out":5536,"duration_ms":52970,"significance":"If the central claim holds, CTL-PINN is a useful and simple extension of existing time-domain-decomposition PINN methods, and the Three Gorges application is a noteworthy real-world demonstration of physics-informed learning with limited measured data. The paper is honest about the main weakness: Conclusion (2) explicitly acknowledges that self-generated supervised points carry error and lead to error accumulation. No error bound or convergence analysis is provided, the numerical comparisons are based on single runs with hand-tuned step sizes, and no code is released. The method is therefore best viewed as an empirically promising training schedule rather than a proven algorithmic improvement. The three examples, especially the reservoir case with independent measured data, give qualitative support to the claim, but the evidence base is not yet strong enough for the paper's broad conclusions.","major_comments":[{"comment":"The distinguishing mechanism of CTL-PINN is the reuse of the previous model's predictions as additional supervised data and as transfer-window initial conditions. The paper itself acknowledges in Conclusion (2) that these pseudo-labels contain error and cause error accumulation, but it provides no bound, contraction estimate, or numerical control showing that the pseudo-label error stays small relative to the residual and data losses over the reported horizons. Consequently, the claimed advantage over standard PINN could in principle be an artifact of favorable step sizes rather than a property of the method. I request per-step reporting of the pseudo-label error against the reference solution, an ablation where pseudo-labels are replaced by exact or independent reference data for the first several windows, or a formal accumulation estimate.","section":"Section 3.1-3.2, Eqs. (13), (18)-(19), Conclusion (2)"},{"comment":"All reported L2 errors appear to come from a single training run for each configuration. The methods use stochastic optimizers (Adam in Section 4.3, L-BFGS initialized randomly in Sections 4.1 and 4.2) and randomly sampled Latin Hypercube points, so the claim that CTL-PINN is 'stable' and 'robust' is not supported without repeated-seed statistics. Please report means and standard deviations over at least several independent runs, or, if the runs are intended to be deterministic, state the seeding and initialization procedure explicitly and justify why the comparison is not run-to-run dependent.","section":"Section 4, Table 3, Figs. 5, 8, 12"},{"comment":"The method's performance depends on the free hyperparameters ΔT_pc, ΔT_ct, the curriculum step count n, and the transfer step count m. The text gives only heuristic guidance (ΔT_pc ≤ T_p/4, T_p < T_c ≤ 3T_p, ΔT_ct ≤ T_c/4) and one configuration per example. No sensitivity analysis is provided, so it is unclear how robust the reported advantage over curriculum-only and transfer-only baselines is to these choices. Please include a systematic variation of step sizes and step counts around the chosen values, and show how the final L2 errors and step counts change.","section":"Algorithm 1, Sections 3.1-3.2"},{"comment":"Section 4.3.1 states that the reservoir model 'does not have initial and boundary conditions,' yet the Saint-Venant system is an initial-boundary-value problem and the subsequent text describes discharge data at segment endpoints and water level at a hydrological station. The mathematical role of these time series is unclear: are endpoint discharges used as boundary conditions, is the hydrological-station water level only an observation, and how is the state initialized at t=0? Please specify exactly which terms enter the loss function (Eqs. (41)-(44)) and how the problem is regularized, since this directly affects the reproducibility and validity of the inversion and 24-hour prediction claims.","section":"Section 4.3.1, Eqs. (34)-(35)"}],"minor_comments":[{"comment":"The naming is inconsistent: the abstract refers to 'CL-PINN' for curriculum learning only, while Section 4.1.1 and Figure 5 use 'CT-PINN' for the same variant; please unify the terminology.","section":"Abstract, Section 4.1.1, Fig. 5"},{"comment":"The displayed source term Q(x,y,z,t) is garbled by missing operators and parentheses, making it difficult to verify that the analytical solution satisfies the governing equation; please re-typeset the equation.","section":"Section 4.1, Eq. (27)"},{"comment":"The initial-condition dataset formula in Eq. (19) is badly formatted and partially unreadable; please rewrite it so that the time at which the source-model predictions are evaluated is clear.","section":"Section 3.2, Eq. (19)"},{"comment":"The text refers to an 'analytical solution' for the Kirchhoff plate, but the captions and text also refer to the 'ABAQUS solution'; please clarify whether ABAQUS is the reference and, if so, state the mesh and solver accuracy used to generate it.","section":"Section 4.2, Figs. 13-14"},{"comment":"The columns 'Error Q' and 'Error Z' are not defined; please specify whether these are relative L2 errors, absolute errors, or another metric, and give the test-set time interval used for the error computation.","section":"Section 4.3, Table 3"},{"comment":"The data availability statement says code for sections other than 4.3 is available 'upon request'; for a computational paper, releasing code and data in a repository would substantially improve reproducibility.","section":"Data availability"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of a computational mechanics or scientific machine learning journal, and the Three Gorges application is attractive. The novelty is incremental relative to existing time-domain-decomposition PINN literature, but I would not reject on that ground. My main concern is evidentiary: the load-bearing pseudo-label mechanism is acknowledged to accumulate error, yet the paper provides no control experiment, error analysis, or repeated-run statistics. These issues should be addressable in a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: this is a solid incremental paper. It takes two existing ideas—curriculum learning and transfer learning—and combines them into a two-phase schedule for time-domain decomposition in PINNs. The method is demonstrated on nonlinear waves, a Kirchhoff plate, and a real-data Three Gorges reservoir model. The contribution is not foundational, but the paper is honest and the empirical comparisons are meaningful.\n\nWhat is new: the specific two-phase schedule (curriculum first, then transfer) and the application to second-order-in-time PDEs and an inverse problem. The paper cites PT-PINN, Penwarden et al., and AT-PINN, so the novelty is the combination, not the ingredients. The Three Gorges example inverts roughness coefficients and makes 24-hour predictions from measured data; that gives the work some practical value.\n\nWhat it does well: the main comparisons use external references (analytical solution, ABAQUS, measured data), so the central claim is not circular. The authors also state plainly that their self-generated supervised points carry error and that error accumulates. That is more transparent than many PINN papers.\n\nSoft spots: the load-bearing assumption is that the pseudo-labeled points from previous steps are accurate enough to guide later steps. The paper gives no error bound or accumulation analysis, and the step-size recommendations are heuristic, with no sensitivity study. There are no repeated runs or error bars, and the code is not released. These are real gaps, but they are common in the PINN literature and do not contradict the empirical claims. The manuscript also has many garbled equations from typesetting; that needs cleanup.\n\nBottom line: the paper deserves a serious referee. If the authors add multiple seeds, release code, and test step-size sensitivity, the empirical case would be much stronger. The error-accumulation concern is acknowledged and is the natural topic for follow-up work, not a reason to reject.","headline":"A credible, incremental extension of time-sliced PINN training, with honest limitations and a useful real-data demo—worth a review, not a breakthrough.","tokens_in":22187,"tokens_out":3286,"would_cite":false,"duration_ms":28644,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Splitting a long simulation into short windows whose predictions feed the next window keeps physics-informed neural networks accurate over long horizons.","keywords":["physics-informed neural networks","curriculum learning","transfer learning","long-term simulation","time-domain decomposition","inverse problems","Three Gorges Reservoir Area","Saint-Venant equations"],"falsifier":"On a benchmark with a known analytical solution, such as the nonlinear annulus wave problem with $T=250$ s, run CTL-PINN and record the $L^2$ error after every curriculum and transfer step. If the per-step error grows superlinearly---for instance, if doubling the number of steps more than doubles the accumulated error---then the recycled predictions are not accurate enough to sustain the claimed long-horizon gain.","tokens_in":21149,"feed_emoji":"⏱️","tokens_out":7566,"duration_ms":62959,"temperature":0.7,"pith_summary":"Standard physics-informed neural networks (PINNs) lose accuracy and stability when asked to solve time-dependent partial differential equations over long horizons. This paper proposes CTL-PINN, a training schedule that decomposes a long simulation into short subproblems, solves the first with a standard PINN, then advances the solution by alternating two strategies: curriculum learning, which extends the trained time window in small increments, and transfer learning, which stacks new time windows onto the learned model. The claim is that this combination yields more accurate and stable long-term solutions than standard PINN, and does so in fewer steps than using either strategy alone. The method is demonstrated on nonlinear wave propagation, the dynamic response of a Kirchhoff plate, and a hydrodynamic model of the Three Gorges Reservoir Area, including inverse estimation of source terms and roughness coefficients.","feed_headline":"Split long simulations into short learned steps to keep PINNs stable","feed_subtitle":"Curriculum and transfer learning recycle earlier predictions as teaching data, holding accuracy where standard PINNs drift.","key_machinery":"The central object is CTL-PINN, a closed training loop whose state is the trained network from the previous time window. What carries the argument is the additional supervised learning dataset $\\tau_{sp}^{(k)}$: each new subproblem is trained not only on the PDE residual, initial, and boundary data, but also on the previous network's predictions at points in the overlap of the old and new time domains (Eqs. (13) and (18), with losses (16) and (22)). Curriculum learning uses these points to widen the window in small steps; transfer learning reuses parameter initialization plus predicted initial conditions to stack non-overlapping windows. The practical rule that curriculum steps stay bounded by $T_p/4$ and transfer steps by $T_{ct}/4$ is what keeps the recycled predictions close enough to the true solution to serve as teaching signals.","core_discovery":"The paper's central claim is that the failure of standard PINNs on long-horizon evolution problems can be overcome by a two-phase training schedule that reuses the previous model as both initialization and teacher. In the curriculum phase, the network trained on $(0,T_p]$ is fine-tuned on successively larger windows $(0,T_p+k\\Delta T_{pc}]$, with predictions of the previous network on the overlap region added as extra supervised points (Eqs. (13) and (16)). In the transfer phase, the model from $(0,T_c]$ is carried forward window by window, using predicted values at the new window's initial time as its initial condition and overlap predictions as teaching data (Eqs. (18), (19), and (22)). The resulting method is reported to keep $L^2$ errors small where standard PINN diverges, to reach a given horizon with fewer training steps than curriculum-only or transfer-only variants, and to extend the same machinery to inverse problems and to field data with limited measurements.","pith_inferences":["The fixed step-size recommendations ($\\Delta T_{pc}\\le T_p/4$ and $\\Delta T_{ct}\\le T_{ct}/4$) are heuristics; a natural extension the paper does not pursue is adaptive step selection driven by the measured mismatch between recycled predictions and any available observations.","If the recycled predictions stay accurate, the same scheme could serve as an online data-assimilation loop for operational river forecasting: each window injects sparse gauge data and corrects drift before the next extrapolation step.","The plate example suggests the schedule also alleviates time-causality violations in continuous PINNs, but the paper does not quantify this; comparing CTL-PINN against causality-weighted loss formulations on the same benchmarks would isolate the mechanism.","A stress test beyond the reported 24-hour forecast horizon, or with rainfall-induced lateral inflow where the model currently sets $q_l=U_l=0$, would show whether the short-term constraint effect that powers extrapolation persists under stronger nonlinearity."],"forward_implications":["For a fixed long horizon, CTL-PINN reaches a stable solution with fewer training steps than curriculum-only (CL-PINN) or transfer-only (TL-PINN) variants, because curriculum learning establishes a larger accurate base window before transfer stacking.","The same schedule handles second-order time derivatives and systems of PDEs, including inverse problems: source-term inversion for nonlinear waves and roughness-coefficient inversion for the reservoir both stay accurate where standard PINN drifts.","Adding real measurements at intermediate times reduces the error accumulation caused by recycled predictions, so the method is directly applicable to data-scarce field problems.","For new problems, the paper's recommended bounds---curriculum step at most $T_p/4$ and transfer step at most $T_{ct}/4$---give a concrete starting configuration.","In the Three Gorges case, the method produces 24-hour-ahead predictions of discharge and water level from hydrological-station data while calibrating the roughness parameter."],"supporting_citations":[{"why":"Provides the standard PINN formulation and optimization settings that CTL-PINN extends and compares against.","marker":"(Raissi et al., 2019)"},{"why":"Supplies the curriculum learning principle that drives the time-window expansion phase.","marker":"(Bengio et al., 2009)"},{"why":"Supplies the transfer learning formalism used for stacking successive time windows.","marker":"(Weiss et al., 2016)"},{"why":"Documents failure modes of standard PINNs on long-duration problems, motivating the time-slicing strategy.","marker":"(Krishnapriyan et al., 2021)"},{"why":"Presents a unified time-domain decomposition framework with overlapping windows that informs the transfer phase.","marker":"(Penwarden et al., 2023)"},{"why":"Introduces PT-PINN, a pre-training and curriculum-like expansion strategy that CTL-PINN builds on and extends.","marker":"(Guo et al., 2023)"},{"why":"Provides the learning-rate annealing algorithm used to adaptively balance loss terms in the reservoir inverse problem.","marker":"(Wang et al., 2020)"},{"why":"Introduces conservative domain-decomposition PINNs, an earlier decomposition approach that motivates splitting the computational domain.","marker":"(Jagtap et al., 2020)"}],"fun_headline_variants":["Break long physics simulations into short learned windows for stable PINNs","Curriculum and transfer learning keep PINNs accurate over long horizons","Reuse past PINN predictions as teaching data to extend simulation time","CTL-PINN: split simulation into short segments to beat standard PINN drift","Long-term PINN stability via curriculum and transfer learning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the previous network's predictions, recycled as additional supervised learning points and as initial conditions for the next window, are accurate enough to guide the next subproblem; if their errors are too large, the errors compound and the claimed advantage over standard PINN disappears, as the paper itself acknowledges when it notes that there is a certain error between the predicted data and the exact values.","fun_headline_variants_meta":{"raw":{"variants":["Break long physics simulations into short learned windows for stable PINNs","Curriculum and transfer learning keep PINNs accurate over long horizons","Reuse past PINN predictions as teaching data to extend simulation time","CTL-PINN: split simulation into short segments to beat standard PINN drift","Long-term PINN stability via curriculum and transfer learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00063,"raw_usage":{"total_tokens":2910,"prompt_tokens":945,"completion_tokens":1965,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":561,"completion_tokens_details":{"reasoning_tokens":1876}},"tokens_in":561,"tokens_out":1965,"duration_ms":12207,"temperature":1.0,"reasoning_tokens":1876,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T13:05:47.203553+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a benchmark with a known analytical solution, such as the nonlinear annulus wave problem with $T=250$ s, run CTL-PINN and record the $L^2$ error after every curriculum and transfer step. If the per-step error grows superlinearly---for instance, if doubling the number of steps more than doubles the accumulated error---then the recycled predictions are not accurate enough to sustain the claimed long-horizon gain.","supporting_citations":[],"review_version":1}