{"id":"b3c59503-1e3d-4a99-b287-59cacc8a27f4","arxiv_id":"2608.00082","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"pyMOR now unifies model-based and data-driven model order reduction in one open Python framework, demonstrated on Navier-Stokes and mass-spring-damper benchmarks.","lead":"This paper describes newly added data-driven model order reduction algorithms in the pyMOR software library, including DMD, Loewner, AAA, ERA, and POD-ML, and evaluates them against classical projection-based methods on flow and control benchmarks. It is a software-consolidation contribution that may lower the barrier to using data-only surrogate models in practice.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Numerical evidence for data-driven MOR is confined to two favorable regimes (high-ν Navier-Stokes, periodic window); transfer to the advertised convection-dominated scope is untested.","rationale":"The reader's weakest_assumption correctly identifies the favorable benchmark regimes as the key threat to external validity. This is the most load-bearing concern because the paper's central claim—while about software scope—is supported by 'practical examples' of the data-driven methods working well. If those examples only work in regimes chosen to avoid known difficulties (slow Kolmogorov N-width, non-periodicity), then the claim that pyMOR is uniquely useful for real-world convection-dominated or non-periodic problems is not established. The paper explicitly self-limits (Sec. 6.1), which is honest, but that limitation means the headline 'extensive selection in a unified framework' is not backed by evidence in the most challenging MOR scenarios. The reader's other issues (DMD fit-quality, ML variance, post-hoc truncation ranks) are real but secondary; they affect individual numbers, not the overall transferability. A concrete extension of the experiment to a low-ν convection-dominated case would settle whether the concern lands. The verdict should remain CONDITIONAL, as the architecture claim is plausible and the code is available, but the numerical demonstrations need broadening or a clear disclaimer about the limited regime.","tokens_in":34699,"tokens_out":19563,"duration_ms":217615,"concrete_test":"Run the POD-ML and POD-DEIM pipelines on the same Navier-Stokes benchmark at ν=10^-4 (or with a non-periodic inflow) and compare max relative state/drag/lift errors (24) and speedups against the ν∈[10^-2,10^0] results. If errors increase by more than an order of magnitude, the demonstrated performance does not transfer to the lower-viscosity regime that the paper explicitly avoids.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Sec. 1) is that pyMOR uniquely integrates model-based and data-driven MOR for both control systems and parameterized PDEs in weak formulation. The demonstration rests on two benchmarks whose structure is explicitly favorable. Sec. 6.1 restricts the parametric Navier-Stokes study to ν∈[10^-2,10^0] \"to avoid challenges due to slow Kolmogorov N-width decay [58]\", and the only DMD test is a periodic vortex-shedding case at ν=10^-3 restricted to t∈[6,8] (Sec. 6.1.4). For convection-dominated problems (low ν, non-periodic dynamics), data-driven methods like POD-ML and DMD are known to degrade because of slow N-width decay and lack of periodicity. The paper acknowledges the restriction but does not provide any experiment outside it. If a user applies the advertised methods to a real convection-dominated problem, the demonstrated accuracy (Tables 1–2, Figs. 10–12) and speedups may not hold. This does not attack the software architecture claim, but it directly undermines the paper's evidence that the unified framework is useful for the broad class of problems implied by the central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents the data-driven model order reduction (MOR) capabilities added to pyMOR, including POD-based methods (POD-ML with VKOGA, DNNs, and GPR), DMD, the Loewner framework, AAA/p-AAA, and ERA. It describes how these methods fit into pyMOR's Operator/Model abstraction and demonstrates them on two test cases: a parametrized incompressible Navier-Stokes problem, where POD-DEIM, POD-ML, a hierarchical POD-DEIM-to-POD-ML pipeline, and DMD are compared; and a mass-spring-damper chain, where model-based (BT, IRKA, SOBTp, PH-IRKA) and data-driven (Loewner, AAA, ERA, p-AAA) system-theoretic methods are compared. The paper claims that pyMOR is the only software library providing such an extensive selection of model-based and data-driven MOR algorithms in a unified framework for both control systems and parameterized PDEs in weak formulation.","tokens_in":35057,"tokens_out":6733,"duration_ms":79208,"significance":"If the framework claim holds, the paper is a valuable contribution to the MOR software ecosystem. It provides an open-source implementation, companion code with a DOI, and extensive numerical experiments spanning both PDE-based and control-system problems. The unified interface-based design, which allows model-based and data-driven methods to be combined in hierarchical pipelines, is a genuine strength. The main caveat is that the numerical evidence for the data-driven methods is obtained in deliberately favorable regimes, and several reporting gaps affect the strength of the stated comparisons. These issues are fixable and do not invalidate the architecture claim, but they are central to the paper's demonstration that the unified framework is practically useful across the advertised scope.","major_comments":[{"comment":"The DMD errors reported in Fig. 10 are reconstruction errors computed on the same time interval [6,8] that is used to build the DMD model. The text states that the solution is 'restricted to the time interval [6,8]', and the DMD code in Section 5.3 fits W and omegas from the same snapshot sequence X that is later compared against. Thus the reported 'relative errors' measure fit quality, not predictive accuracy. This is especially consequential because the eigenvalues are shown to lie on the unit circle, confirming that the test case is periodic. Please relabel these as training/reconstruction errors, or preferably evaluate on a holdout interval (e.g., t in [8,10]) or an unseen initial condition and report generalization error.","section":"Section 6.1.4, Fig. 10"},{"comment":"The machine learning results in Table 2 are single runs. The NeuralNetworkRegressor uses random initialization, mini-batching, and early stopping, so the reported errors and the relative ranking of VKOGA, DNN, and GPR may depend on the random seed. The statement 'restarting the training with different random initial conditions is not necessary' is not a substitute for reporting variance. Please provide means and standard deviations over at least 5--10 independent runs, or fixed seeds for all methods, and state the number of runs. This is needed to support conclusions such as the DNN benefiting from 200 training points in the POD-DEIM-ML setting.","section":"Section 6.1.3, Table 2"},{"comment":"The numerical evidence for data-driven MOR is confined to two favorable regimes. The parametric Navier-Stokes study restricts the viscosity to [10^-2, 10^0] explicitly to avoid slow Kolmogorov N-width decay [58], and the only DMD experiment uses a periodic vortex-shedding regime at nu=10^-3 on t in [6,8]. These restrictions are acknowledged in the text, but the abstract and central claim (Sec. 1) advertise data-driven methods for a broad class of parameterized PDE problems. As presented, the demonstrated accuracy and speedups for POD-ML and DMD may not transfer to advection-dominated, non-periodic, or transport-dominated problems. Please either add at least one experiment outside these favorable regimes, or explicitly scope the claims in the abstract and conclusion to the tested parameter and time ranges.","section":"Section 6.1"},{"comment":"The hierarchical POD-DEIM-ML pipeline is presented as enabling efficient training-data generation, but the numerical results show a mixed picture. For VKOGA and GPR, the POD-DEIM-ML state errors (2.41e-04) are an order of magnitude larger than the corresponding POD-ML errors (3.21e-05), and the speedups are comparable. The paper notes the state-error increase but does not discuss its structural reason: the hierarchical ROM inherits the error of its teacher ROM, so the extra training data did not compensate for the reduced fidelity of the teacher in the kernel-based methods. Please add a discussion of this accuracy/cost trade-off, since it directly bears on the claim that pyMOR enables 'flexible and efficient hierarchical model reduction pipelines.'","section":"Section 6.1.3, Table 2"}],"minor_comments":[{"comment":"The heading contains a typo: 'algortihms' should be 'algorithms'.","section":"Section 5.2.1"},{"comment":"There is a typo in the first sentence: 'recoverd form' should be 'recovered from'.","section":"Section 5.6"},{"comment":"The custom IMEXModel assumes that E, F_i, and g are time-independent, as noted in the code comment, but this assumption is not stated in the main text. Please make it explicit as a limitation of the current implementation.","section":"Section 3.3"},{"comment":"The claim that pyMOR is 'the only software library' with this combination is supported only by a narrative survey and 'to the best of our knowledge.' A feature-comparison table against the cited libraries (e.g., libROM, PyDMD, MORLAB, Pressio) would make the claim easier to verify.","section":"Section 1"},{"comment":"The visual comparison between the FOM and DMD solutions at the final time is qualitative. Since the DMD errors in Fig. 10 are reconstruction errors on the training interval, the figure caption should specify that this is a reconstruction, not a forecast.","section":"Section 6.1.4, Fig. 12"}],"recommendation":"major_revision","confidential_remarks":"The paper is substantive and the software contribution is real, but the numerical evaluation needs work before publication. The DMD evaluation is circular as reported, the ML comparisons lack variance reporting, and the scope of the numerical evidence is narrower than the abstract implies. These are fixable within the manuscript's scope, so I recommend major revision rather than rejection. The uniqueness claim in Sec. 1 would also benefit from a systematic comparison table rather than a narrative survey."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this is a software paper, not an algorithms paper. The new content is the integration of data-driven MOR methods (Loewner, AAA, p-AAA, TF-IRKA, DMD, ERA, POD-ML) into pyMOR's existing model-based framework, plus a comparative numerical study. The architecture is genuinely well done: the VectorArray/Operator/Model abstractions let you mix model-based and data-driven workflows in one tool, and the hierarchical pipeline where a POD-DEIM ROM generates training data for a POD-ML ROM is a nice concrete demonstration. Code and data are available with a DOI, which is real evidence.\n\nThe main soft spots are in the numerical reporting, not the software. First, the DMD errors in Fig. 10 are reconstruction errors on the same time window used to fit the modes. That's fit quality, not prediction, and it overstates what DMD delivers. Second, the ML results in Table 2 are single runs with no seeds or variance; the VKOGA-vs-DNN comparison is shaky when you can't see how much the DNN training randomness matters. Third, the parametric BT/IRKA truncation ranks are chosen post hoc on the test data — minor, but should be stated. Fourth, and this is the stress-test concern that holds up: the Navier-Stokes study is explicitly restricted to ν ∈ [10^-2, 10^0] to avoid slow Kolmogorov N-width decay, and the DMD example is a periodic window. So the advertised convection-dominated scope is untested. The paper acknowledges the restriction, but the performance claims in the abstract should be tempered accordingly. These are addressable reporting gaps, not load-bearing errors.\n\nThe central claim — that pyMOR is uniquely positioned as a unified framework for both model-based and data-driven MOR — looks plausible from the survey and the architecture description. I'd be comfortable citing the software and the framework, less comfortable citing the speedups without the caveats.\n\nThis paper deserves serious peer review. The referees should ask for seeds/variance on ML, a predictive-error evaluation for DMD, and a more careful statement of scope. I'd send it to review with that expectation.","headline":"A solid, thorough software-integration report for pyMOR's data-driven MOR suite; the architecture and code are the contribution, but the numerical comparisons have reporting gaps that need fixing before you rely on the performance claims.","tokens_in":35546,"tokens_out":1697,"would_cite":true,"duration_ms":22467,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["41A20","65-04","65D15","65M60","65N30","93C15","93C20"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that one open-source Python library, pyMOR, is the only MOR framework that unifies an extensive set of model-based and data-driven reduction algorithms for both parametric PDEs and control systems, and supports hybrid pipel","keywords":["pyMOR","model order reduction","data-driven reduction","parametric PDEs","LTI systems","dynamic mode decomposition","Loewner framework","surrogate modeling"],"falsifier":"Run the POD-ML and DMD pipelines on a convection-dominated parametric flow with viscosity below 1e-2, or on a non-periodic unsteady flow, and compare against the reported relative errors; if kernel surrogates degrade below roughly 1e-3 or DMD requires far more than five modes, the practical breadth claim is weakened. A survey-level check also settles the uniqueness claim: any actively maintained open-source package that offers transfer-function interpolation, state-space identification, and parametric projection methods through one model interface would directly refute the 'only library' asser","tokens_in":34520,"feed_emoji":"🐍","tokens_out":7026,"duration_ms":90029,"temperature":0.7,"pith_summary":"This paper describes recent additions of data-driven model order reduction (MOR) algorithms to the open-source library pyMOR and argues that pyMOR is the only software library offering such an extensive selection of model-based and data-driven MOR algorithms in a unified framework. It demonstrates that a user can expose a full-order model once through pyMOR's abstract interfaces and then choose among projection-based, data-only, and hybrid strategies. In the Navier-Stokes benchmark, projection-based ROMs run 47-74x faster than the full-order model, kernel-based machine-learning surrogates reach speedups around 1.8 million for direct output mapping, and using a model-based ROM to generate training data for a data-driven ROM cuts snapshot costs from about 9.4 hours to about 1.2 hours and improves neural-network accuracy. On a mass-spring-damper control system, the paper compares data-driven transfer-function methods (Loewner, AAA, ERA) with model-based balanced truncation and IRKA, finding the data-driven ROMs about an order of magnitude less accurate at the same reduced order.","feed_headline":"One library unifies physics- and data-driven surrogates","feed_subtitle":"Swap between projection-based ROMs and data-only surrogates, and chain them, without changing solver wrappers.","key_machinery":"The organizing mechanism is pyMOR's interface layer: all equations appear as Model objects built from Operator and VectorArray objects, so projection-based reductors and data-driven reductors consume the same data structures and a model-based ROM can be reused as the data source for a data-driven ROM. Supporting machinery includes HAPOD for memory-limited snapshot compression (a three-level POD tree in the Navier-Stokes example), DEIM with least-squares residual minimization for hyperreduction, kernel interpolation (VKOGA) for low-data parameter-to-coefficient maps, and transfer-function interpolation via Loewner matrices, barycentric AAA, and ERA from impulse-response data.","core_discovery":"The paper's central claim is that a single software framework can meaningfully span both classical model-based model order reduction and recent data-driven approaches by expressing every full-order model through the same abstract Operator, VectorArray, and Model interfaces. It establishes this by adding data-driven reductors for POD-ML, DMD, Loewner, AAA, and ERA to the existing projection-based toolbox, and by showing a hybrid pipeline in which a projection-based POD-DEIM ROM generates training data for a kernel- or neural-network surrogate. The numerical evidence is presented as supporting evidence: model-based ROMs reach 47-74x speedups with relative state errors around 1e-4, kernel-based","pith_inferences":["Editorial inference: the same interface abstraction that enables model/data hybrid pipelines also enables per-parameter method selection—for example, using DMD in periodic parameter regimes and POD-ML in transient regimes—which the paper mentions as future work through an adaptive hierarchy.","Editorial inference: the VKOGA advantage over neural networks at 30 training parameters suggests that greedy kernel centers may remain the data-efficient choice as parameter dimension grows, but the paper's parametric control example is only two-dimensional and does not test this directly.","Editorial inference: because a single 71,352-DoF Navier-Stokes trajectory already exceeds 8 GB of memory, HAPOD-style hierarchical compression is likely required for any high-dimensional unsteady output; whether the same speedups hold after accounting for external-solver I/O and wrapper overhead is not measured in the paper."],"forward_implications":["With a full-order model exposed through pyMOR's interfaces, comparing projection, data-only, and hybrid reduction methods becomes a matter of selecting a reductor rather than re-implementing a workflow.","A model-based ROM can act as a cheap surrogate for generating training data for a data-driven ROM: the paper shows 200 training parameters generated from the POD-DEIM ROM cost about 1.2 hours versus about 9.4 hours from the full-order model, and the neural-network surrogate improves by almost an order of magnitude.","Interpolation-based kernels (VKOGA, GPR) are the better default in low-data regimes; deep neural networks catch up only when hundreds of training parameters are available.","For periodic unsteady flows, five DMD modes reproduce vortex-shedding velocity fields to visual accuracy, with error decaying as the number of modes increases.","For LTI control systems, data-driven interpolation methods (Loewner, AAA, ERA) produce ROMs within roughly one order of magnitude of model-based balanced truncation and IRKA at reduced order 20, and p-AAA matches parametric BT/IRKA to about 1e-2 error on the tested damping range."],"fun_headline_variants":["pyMOR unifies physics- and data-driven model reduction","One library bridges projection-based and data-only surrogates","PyMOR adds DMD, Loewner, and neural-net reducers to its toolbox","Build hybrid ROM pipelines with pyMOR's unified interfaces","From PDE solver to data-only surrogate: pyMOR does both"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The demonstration depends on the favorable structure of the benchmarks: the Navier-Stokes parameter range is restricted to high viscosities (1e-2 to 1) to avoid slow Kolmogorov N-width decay, and DMD is tested only in a periodic vortex-shedding window; if the advertised real-world problems are not similarly compressible or periodic, the shown accuracy and speedups are not guaranteed to carry over.","fun_headline_variants_meta":{"raw":{"variants":["pyMOR unifies physics- and data-driven model reduction","One library bridges projection-based and data-only surrogates","PyMOR adds DMD, Loewner, and neural-net reducers to its toolbox","Build hybrid ROM pipelines with pyMOR's unified interfaces","From PDE solver to data-only surrogate: pyMOR does both"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00025,"raw_usage":{"total_tokens":1360,"prompt_tokens":683,"completion_tokens":677,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":427,"completion_tokens_details":{"reasoning_tokens":595}},"tokens_in":427,"tokens_out":677,"duration_ms":39673,"temperature":1.0,"reasoning_tokens":595,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T01:20:13.524758+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the POD-ML and DMD pipelines on a convection-dominated parametric flow with viscosity below 1e-2, or on a non-periodic unsteady flow, and compare against the reported relative errors; if kernel surrogates degrade below roughly 1e-3 or DMD requires far more than five modes, the practical breadth claim is weakened. A survey-level check also settles the uniqueness claim: any actively maintained open-source package that offers transfer-function interpolation, state-space identification, and parametric projection methods through one model interface would directly refute the 'only library' asser","supporting_citations":[],"review_version":1}