{"id":"4519cd51-fd94-401e-b476-7cd1e0307f5d","arxiv_id":"2411.19125","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"PIDO trains an auto-decoded latent-space dynamics model with physics-informed losses and two regularizers, reporting lower errors on unseen initial conditions, coefficients, and time horizons than PI-DeepONet, PINODE, and MAD.","lead":"This paper introduces PIDO, a physics-informed neural network that learns compact latent representations of PDE solutions and their dynamics, aiming to generalize to new initial conditions, coefficients, and future times without needing simulated trajectory data. It reports large error reductions over existing solvers on Burgers, KdV, and 2D Navier-Stokes benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The core claim that PDE-residual training alone identifies the true latent dynamics is not secured: the objective admits a trivial branch where D(c_t) is the zero solution for t>0, and the printed regularizer in Eq.","rationale":"The reader's weakest assumption correctly identifies that the physics-informed residual plus the two regularizers is only empirically shown to yield the true latent dynamics. My stress test points to a concrete failure mode the paper does not rule out: because D(c_t) is free for t>0 and the benchmark PDEs admit the zero solution, the objective can, in principle, be nearly minimized by a latent trajectory that leaves the initial-condition reconstruction at t=0 and collapses to a zero-output decoder state at sampled later times. This is precisely the 'trivial or non-physical solution' the reader flagged, and the paper's diagnostic in Section III-D relies on visualization rather than a formal argument. The additional concrete issue with Eq. (14) strengthens the concern and supports the conditional verdict rather than a higher one. I did not find a basis to reject the paper outright: the method is plausible, the comparisons are broad, and the failure mode I describe may be empirically prevented by the smoothing and alignment regularizers. However, the absence of code, error bars, or an identifiability argument, together with the mis-printed regularizer, means the central claim remains conditionally supported. The verdict should stay CONDITIONAL as the reader recommended.","tokens_in":18960,"tokens_out":6986,"duration_ms":87927,"concrete_test":"Retrain PIDO on CE1 exactly as specified, and at evaluation sample the PDE residual on a ten-times-finer temporal grid than the 60 training steps (e.g., 600 equally spaced points in [0,1.92]) for the 512 test initial conditions. Compare the fine-grid PDE residual with the coarse-grid residual, and the L2 error separately over the training horizon and the extrapolation horizon. If the fine-grid residual is substantially larger than the coarse-grid value while the t>0 solution error remains low, the physics loss was only small on the training discretization, and a trivial or non-physical latent branch cannot be excluded.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing premise is that minimizing lPDE over unrolled latent trajectories drives F and D to the true solution operator without interior data. Equation (13) supervises D(c_t) through lIC only at t=0 and through lPDE/lBC at discrete t>0. For the benchmark PDEs, the zero field is itself a solution with zero residual and zero boundary loss. A continuous Neural-ODE trajectory c_t can start from the c0 that reconstructs the initial condition and then move to a decoder state whose output is (near) zero at all sampled t>0; if the transition occurs between collocation points, the discrete residual loss is insensitive to it. RA does not remove this degeneracy because it aligns c_t to auto-decoded pseudo-labels of the model's own predictions, and RS is a generic smoothness penalty. The paper provides no identifiability or convergence argument, only the visualization in Section III-D. Moreover, Eq. (14) is mathematically incorrect as printed: for a Gaussian epsilon, E[||epsilon^T J epsilon||^2] equals 2||J||_F^2 + (tr J)^2, not ||J||_F^2, so the reported smoothing regularizer is not fully specified. Without released code or error bars, one cannot verify whether the implemented RS is the printed one or whether the reported errors reflect this unidentified objective.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PIDO, a physics-informed neural PDE solver that combines auto-decoding of spatial fields into a latent space with a coefficient-conditioned Neural ODE for latent dynamics. Training minimizes initial-condition reconstruction, PDE residual, and boundary losses on unrolled latent trajectories, augmented by two latent-space regularizers: Latent Dynamics Smoothing (RS) and Latent Dynamics Alignment (RA). The authors evaluate on 1D combined equations (CE1-CE3) and 2D Navier-Stokes equations (NS1-NS2) with held-out initial conditions, coefficients, and temporal horizons, reporting large L2-relative-error margins over PI-DeepONet, PINODE, and MAD, as well as favorable comparisons with data-driven baselines. They also demonstrate transfer of the learned representations to long-term integration and inverse problems.","tokens_in":19348,"tokens_out":7448,"duration_ms":70590,"significance":"If the reported results are reproducible, PIDO is a meaningful advance: it is the only method among those compared that generalizes across initial conditions, PDE coefficients, and time horizon in a residual-only (data-free of exact solutions) setting. The held-out evaluation protocol is appropriate, and the reported margins are large, with ablations in Table IV showing that both regularizers contribute to the final performance. The downstream-task results in Table V provide additional evidence of representation transfer. The main weaknesses are that the printed definition of RS in Eq. (14) contains an incorrect identity, the latent-dynamics objective is only heuristically justified with no identifiability or convergence analysis, and no code or error bars are provided. These issues need to be addressed before the claimed generalization can be taken as fully established.","major_comments":[{"comment":"The second equality in Eq. (14) is not an identity: for J = ∇F(c_t, α), E_{ε∼N(0,I)}∥ε^T J ε∥² = (tr J)² + 2∥(J+J^T)/2∥_F², which is not equal to ∥J∥_F² in general. As printed, the regularizer is therefore underspecified: implementing the quadratic-form expectation gives a different loss from implementing the Frobenius norm. Please state the exact loss used in the experiments, correct Eq. (14), and clarify whether the implementation penalizes the true Jacobian norm or the quadratic-form expectation.","section":"Section III-D, Eq. (14)"},{"comment":"The training objective in Eq. (13) admits a trivial-solution branch: because the zero field is a solution of the benchmark PDEs with zero boundary loss, a trajectory that reconstructs φ at t = 0 and then drives D(c_t) to zero at the discrete collocation times t > 0 can make the residual and boundary terms vanish without approximating the true dynamics. Neither RS nor RA excludes this branch: RS penalizes F only at sampled times, and RA aligns c_t to an auto-decoding of the model's own output. The paper provides only qualitative visualizations in Section III-D rather than a quantitative argument. Please add a concrete diagnostic, such as evaluating D(c_t) at 5–10× finer temporal resolution than the training collocation grid for several test trajectories, and report sensitivity of the final errors to the residual time step and to the number of auto-decoding steps. This would demonstrate that the reported generalization is not an artifact of the discrete time grid.","section":"Section III-C, Eq. (13); Section III-D"},{"comment":"RA is self-referential: the anchor c̃_t is produced by auto-decoding D(c_t), so the regularizer couples c_t to a pseudo-label that is a function of the same prediction. The claim that c̃_t 'does not exhibit drift' is supported only by one qualitative example (Figure 3), and the mechanism could in principle be a generic shrinkage of c_t toward a low-norm or low-variance region. Please provide quantitative statistics of c_t and c̃_t across the extrapolation horizon, such as distances to the training embedding distribution, and compare RA against simpler alternatives (a fixed pretrained encoder, a norm penalty, or a stop-gradient target) to isolate the effect of the proposed anchor mechanism.","section":"Section III-D.2, Fig. 4"},{"comment":"All reported errors come from a single run and no code is released. Physics-informed training is known to be sensitive to initialization and loss weights, so the reported margins, especially the large Out-t gains, need multiple-seed results (mean ± std) for the main comparisons. Without error bars, one cannot assess whether the improvements are statistically robust; without code, the discrepancy between the printed Eq. (14) and the actual implementation cannot be resolved. Please provide at least three seeds per setting for Tables II–V and release the code or a detailed configuration.","section":"Tables II–V; Appendix C"}],"minor_comments":[{"comment":"The notation ε ∼ N(0,1) should be ε ∼ N(0,I) when ε is a vector; as written, the quadratic form is dimensionally inconsistent.","section":"Eq. (14)"},{"comment":"The sampling line uses {c̄_t} while the initialization and the alignment update use {c̃_t}; please unify the notation for the anchor embeddings.","section":"Algorithm 1"},{"comment":"The notation c_t(x) in the chain rule is misleading because c_t is a latent vector independent of x; the term should be written as ∂ũ_t(x)/∂c_t · F(c_t, α).","section":"Eq. (10)"},{"comment":"The symbol E is used both for the encoder and for expectation; consider using a different notation for the encoder to avoid confusion.","section":"Section III-B, Eq. (6)"},{"comment":"The number S of test-time auto-decoding steps in Algorithm 2 is not reported; please state S for each benchmark along with the collocation-point sampling scheme.","section":"Appendix C-A; Algorithm 2"},{"comment":"The ablation of RS and RA is reported only for the NS2 scenario; reporting the same ablations on at least one 1D scenario would strengthen the claim that both regularizers are generally necessary.","section":"Table IV"},{"comment":"Computational cost is reported only for the CE1 setting; a brief statement for NS1/NS2 would help assess scalability.","section":"Table X"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the journal's scope and the empirical results are promising, but the main risk is that the central regularizer is mis-specified as printed and the empirical claims rest on single runs without code. I would encourage the editor to require corrected equations, error bars, and code or detailed configuration before acceptance. The authors already acknowledge the periodic-boundary-only limitation in the discussion, which is appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The two things to know: this is a genuine attempt at a data-free PDE solver that generalizes over initial conditions, coefficients, and time horizons in one model; and the benchmark numbers, if they hold up, are a real step forward for the latent-dynamics-plus-PINN line. But as printed, the paper has a concrete mathematical mistake in its key regularizer, and it does not secure its central claim that physics-informed residual training alone identifies the true latent dynamics.\n\nWhat's actually new: the combination of auto-decoded implicit neural representations, coefficient-conditioned Neural ODE dynamics, and physics-informed loss with the two regularizers. The diagnosis of latent-space failure modes—overly complex dynamics causing instability, and embedding drift hurting extrapolation—is a nice addition, and the ablations show both regularizers matter. The evaluation spans Burgers, KdV, combined 1D equations, and 2D Navier-Stokes, with held-out test sets and downstream tasks. That is more than most PINN papers do.\n\nWhere the soft spots are. First, Eq. (14) is wrong as printed. The squared Frobenius norm of the Jacobian is not E_epsilon ||epsilon^T J epsilon||^2 for Gaussian epsilon; that expectation equals 2||J||_F^2 + (tr J)^2. So the regularizer is not fully specified. Since this regularizer is crucial for training stability, the error needs to be fixed and the implemented form stated. Minor related issue: Algorithm 2's auto-decoding step count S is never specified.\n\nSecond, the identifiability concern is real, though I would frame it as a gap rather than a demonstrated failure. The objective in Eq. (13) supervises D(c_t) through the initial condition at t=0 and through residual/boundary losses at discrete t. There is a plausible trivial branch where the trajectory moves to a decoder state that is near zero at all sampled t>0, and the discrete residual loss would not see it. Alignment is a self-consistency regularizer, not a cure for this degeneracy, and smoothing penalizes rough dynamics but does not add information. The paper gives visualizations, no convergence or identifiability argument. The empirical results are consistent with genuine learned dynamics, so I would ask the authors to address the gap explicitly rather than reject the paper.\n\nThird, no code or error bars. Combined with the Eq. (14) issue, I cannot verify the reported losses are what was trained. This is a reproducibility burden the authors should lift.\n\nThe paper is for people working on physics-informed learning and neural operators, especially those interested in temporal extrapolation and parametric PDEs. The central idea is plausible and the experiments are extensive; the problems are fixable. I would send it to serious peer review with the requirement that the math error be corrected, code released, and the identifiability concern discussed. As published, I would not rely on the numbers yet, but it deserves referee time.","headline":"Strong empirical proposal for a data-free latent-dynamics PDE solver, but a concrete math error in its main regularizer and an unproven identifiability claim mean the numbers should be treated with caution until fixed and reproduced.","tokens_in":19799,"tokens_out":3462,"would_cite":false,"duration_ms":40711,"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":"PIDO claims that a physics-informed neural PDE solver can generalize across unseen initial conditions, PDE coefficients, and time horizons by auto-decoding solutions into a latent space and learning latent dynamics with a Neural ODE…","keywords":["physics-informed neural networks","latent space regularization","partial differential equations","spatial-temporal dynamics modeling","neural ordinary differential equations","auto-decoding","temporal extrapolation","Navier-Stokes equations"],"falsifier":"Train the CE1 or NS1 scenario with auto-decoding replaced by a directly trained encoder that outputs $c_0$, keeping the PDE-residual loss and both regularizers unchanged: if In-t and Out-t errors stay at the reported levels, generalization to unseen initial conditions does not depend on the claimed auto-decoding mechanism. A second decisive check is to compare the code’s actual smoothing regularizer with Eq. (14), since the printed equality between $\\|\\nabla F\\|_F^2$ and $E_{\\varepsilon\\sim\\mathcal{N}(0,I)}(\\varepsilon^T \\nabla F \\varepsilon)^2$ is not an identity; re-running Table IV with the printed objective would show which form produces the reported gains.","tokens_in":18765,"feed_emoji":"🧮","tokens_out":12222,"duration_ms":102454,"temperature":0.7,"pith_summary":"PIDO is a physics-informed neural PDE solver built to generalize across configurations of the same PDE family: unseen initial conditions, unseen coefficients, and time horizons beyond those used in training. The paper’s central claim is that this can be achieved without any exact-solution data, using auto-decoding to project solutions into a low-dimensional latent space, a Neural ODE conditioned on the PDE coefficients to evolve the latent codes, and two latent-space regularizers to keep that evolution stable and drift-free. On five benchmark scenarios (1D Burgers, KdV, combined, and 2D Navier-Stokes), PIDO reports the lowest test-set $L^2$ relative errors, including temporal extrapolation errors far below the best baseline, for example 2.24% versus 9.32% on CE1 Out-t. If the claim holds, physics-informed solvers could serve whole families of problems rather than a single problem instance, and do so with far less data than neural operator methods typically require.","feed_headline":"Latent dynamics push physics-informed nets past their training horizon","feed_subtitle":"Same model handles unseen initial conditions and coefficients with no exact solution data.","key_machinery":"The load-bearing object is the pair $(D, F)$ acting on a latent code $c_t$: $D$ is a grid-independent implicit neural representation mapping $(c_t, x)$ to the solution value at $x$, and $F$ is a Neural ODE mapping $(c_t, \\alpha)$ to the time derivative of the latent code, so predictions come from integrating $c_0 = E(\\phi)$ forward and decoding, where $E(\\phi)$ denotes the auto-decoded latent fit to the initial condition. Auto-decoding replaces a trained encoder with a short gradient-descent fit of $c_0$ to the initial condition through $D$, which is what lets the model adapt to unseen initial conditions at test time. The argument is carried by two regularizers in latent space: Latent Dynamics Smoothing penalizes $\\|F(c_t,\\alpha)\\|^2$ and the Jacobian norm $\\|\\nabla F\\|_F^2$ to prevent the physics-informed loss from selecting wildly fluctuating dynamics, and Latent Dynamics Alignment penalizes $\\|c_t - \\tilde{c}_t\\|^2$ where $\\tilde{c}_t$ is obtained by re-encoding the predicted solution through the decoder, keeping unrolled embeddings inside the decoder’s well-behaved range. Together these regularizers are what transform the residual loss from an unstable, non-extrapolating objective into a trainable one.","core_discovery":"The central discovery is that the PDE residual alone, evaluated on unrolled latent trajectories, can supervise a correct latent dynamics model provided those trajectories are regularized in latent space. PIDO encodes each initial condition by auto-decoding (fitting a latent vector with gradient descent through the decoder), integrates that vector with a Neural ODE whose right-hand side is conditioned on the coefficients $\\alpha$, and decodes the result with an implicit neural representation to produce the solution at any spatial point. Training uses the initial-condition reconstruction loss plus the PDE residual and boundary loss, with no exact solution data. The paper identifies two failure modes in the latent trajectories—overly complex dynamics that destabilize training and latent embedding drift that degrades extrapolation—and counteracts them with Latent Dynamics Smoothing and Latent Dynamics Alignment. With both regularizers, PIDO reports test $L^2$ relative errors of 1.48% (CE1 In-t), 3.02% (CE2 In-t), 2.35% (NS1 In-t), and 4.59% (NS2 In-t), with Out-t errors of 2.24%, 7.15%, 8.08%, 5.43%, and 10.02%, respectively, beating PI-DeepONet, PINODE, and MAD in every setting.","pith_inferences":["Editorial extension: because the smoothing regularizer penalizes the Jacobian norm of the dynamics, PIDO should be at its best on smooth, diffusion-dominated solutions; a natural test is to apply the same training recipe to shock-forming equations and anneal the smoothing weight over training.","Editorial extension: the alignment regularizer treats the decoder as a pseudo-inverse, so its effectiveness should degrade if the decoder is made deliberately non-injective (for example by increasing latent dimension); this yields a testable prediction about when drift control will or will not help.","Editorial extension: the same latent-space diagnosis could be applied to other physics-informed operator learners: constraining predicted latent trajectories to the decoder’s well-posed region may be a general cure for the optimization pathology of residual losses, not just for this architecture.","Editorial extension: the reported gains are all on periodic-boundary benchmarks, and the paper notes this limitation itself; the most informative next experiment is a non-periodic boundary or irregular geometry, where the grid-independent decoder should be an advantage if the mechanism is general."],"forward_implications":["A single PIDO model trained on one benchmark family can be applied to unseen initial conditions without retraining; on CE1, CE2, and NS1 the In-t test error drops by 63%, 72%, and 76% relative to the second-best baseline.","The same model predicts beyond its training horizon: the Out-t L2 errors are 2.24% (CE1), 7.15% (CE2), 8.08% (CE3), 5.43% (NS1), and 10.02% (NS2), where every baseline is at least roughly twice as large.","Because training uses only the PDE residual, initial conditions, and boundary conditions, the method does not need exact solution data and is more sample-efficient than data-driven EDM variants such as DINO.","The learned latent representations transfer to downstream tasks: after fine-tuning, a pre-trained PIDO cuts accumulated long-term integration error by 77% and recovers PDE coefficients from as few as two snapshots far more accurately than a from-scratch PINN.","The method also extrapolates outside the training range of Reynolds numbers in NS2, with lower errors than MAD at both low (alpha=550) and high (alpha=1850) unseen values."],"supporting_citations":[{"why":"Supplies the auto-decoding mechanism: embeddings are fit by gradient descent rather than by a trained encoder.","marker":"[12]"},{"why":"Provides the Neural ODE formulation used for the latent dynamics model F.","marker":"[15]"},{"why":"Defines the explicit latent dynamics modeling paradigm (DINO) that PIDO makes physics-informed and compares against.","marker":"[13]"},{"why":"Introduces physics-informed DeepONet, the main operator-learning baseline and a reference for physics-informed training without data.","marker":"[11]"},{"why":"Introduces Meta-Auto-Decoder, the prior auto-decoding PDE solver used as a baseline.","marker":"[18]"},{"why":"Introduces PINODE, the prior physics-informed Neural ODE baseline, and is analyzed for its input-distribution assumptions.","marker":"[19]"},{"why":"Supplies the Jacobian-regularization idea behind Latent Dynamics Smoothing.","marker":"[39]"},{"why":"Provides the long-time integration setup and auto-regressive evaluation strategy used in the downstream task.","marker":"[40]"}],"fun_headline_variants":["Latent-space regularization lets PINNs generalize beyond training","Auto-decoded latent dynamics yield PDE solvers that extrapolate","No exact data: PIDO learns latent physics for unseen settings","Regularized latent trajectories make PINNs extrapolate in time"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is empirical: minimizing the PDE residual over unrolled latent trajectories, together with the two hand-weighted regularizers, drives the dynamics network to learn the true latent evolution rather than a trivial or non-physical solution, and the paper gives no identifiability or convergence argument for this; the printed form of the smoothing regularizer also appears to rely on an equality that is not true in general.","fun_headline_variants_meta":{"raw":{"variants":["Latent-space regularization lets PINNs generalize beyond training","Auto-decoded latent dynamics yield PDE solvers that extrapolate","No exact data: PIDO learns latent physics for unseen settings","Regularized latent trajectories make PINNs extrapolate in time"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0012,"raw_usage":{"total_tokens":4980,"prompt_tokens":1013,"completion_tokens":3967,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":629,"completion_tokens_details":{"reasoning_tokens":3899}},"tokens_in":629,"tokens_out":3967,"duration_ms":44934,"temperature":1.0,"reasoning_tokens":3899,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:30:56.812707+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the CE1 or NS1 scenario with auto-decoding replaced by a directly trained encoder that outputs $c_0$, keeping the PDE-residual loss and both regularizers unchanged: if In-t and Out-t errors stay at the reported levels, generalization to unseen initial conditions does not depend on the claimed auto-decoding mechanism. A second decisive check is to compare the code’s actual smoothing regularizer with Eq. (14), since the printed equality between $\\|\\nabla F\\|_F^2$ and $E_{\\varepsilon\\sim\\mathcal{N}(0,I)}(\\varepsilon^T \\nabla F \\varepsilon)^2$ is not an identity; re-running Table IV with the printed objective would show which form produces the reported gains.","supporting_citations":[{"cited_title":"Neu- ral ordinary differential equations,","cited_arxiv_id":null,"evidence_quote":"Provides the Neural ODE formulation used for the latent dynamics model F."},{"cited_title":"Con- tinuous pde dynamics forecasting with implicit neural representations,","cited_arxiv_id":null,"evidence_quote":"Defines the explicit latent dynamics modeling paradigm (DINO) that PIDO makes physics-informed and compares against."},{"cited_title":"Learning the solution operator of parametric partial differential equations with physics-informed deeponets,","cited_arxiv_id":null,"evidence_quote":"Introduces physics-informed DeepONet, the main operator-learning baseline and a reference for physics-informed training without data."},{"cited_title":"Meta-auto-decoder for solving parametric partial differential equations,","cited_arxiv_id":null,"evidence_quote":"Introduces Meta-Auto-Decoder, the prior auto-decoding PDE solver used as a baseline."},{"cited_title":"Physics-informed neural ode (pinode): embedding physics into models using collocation points,","cited_arxiv_id":null,"evidence_quote":"Introduces PINODE, the prior physics-informed Neural ODE baseline, and is analyzed for its input-distribution assumptions."},{"cited_title":"How to train your neural ode: the world of jacobian and kinetic regularization,","cited_arxiv_id":null,"evidence_quote":"Supplies the Jacobian-regularization idea behind Latent Dynamics Smoothing."},{"cited_title":"Long-time integration of parametric evolution equations with physics-informed deeponets,","cited_arxiv_id":null,"evidence_quote":"Provides the long-time integration setup and auto-regressive evaluation strategy used in the downstream task."}],"review_version":1}