{"id":"e94f6521-2b6b-470c-8a08-43bbfe5f6630","arxiv_id":"2411.08750","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Optimal transport displacement interpolation augments sparse simulation data and maps virtual time to real time, enabling reduced-order predictions for nonlinear advection-dominated flows.","lead":"This paper uses optimal transport to create synthetic intermediate snapshots between stored simulation states, then adds a learned correction to build a reduced-order model of nonlinear atmospheric flows. The approach targets situations where full simulation data are scarce and flow features move sharply, which breaks standard linear compression methods.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"OT displacement interpolation assumes checkpoint-to-checkpoint dynamics are close to a Wasserstein geodesic; for vortical/diffusive evolution the OT plan is not the physical flow map, so synthetic snapshots may be artifacts and the claimed accuracy is not robust.","rationale":"Reader's weakest assumption matches the main risk. The novel component of the method is OT displacement interpolation; if it does not track the physical transport, the data augmentation produces unphysical snapshots and the time-mapping plus residual correction cannot rescue the claim for arbitrary-times. The paper's own Figure 4 concedes that with two checkpoints the interpolated trajectory is of limited physical relevance, and the method converges only as Nc increases. This is consistent with the transport assumption being valid only in the limit of very short intervals, at which point the benefit over simpler interpolation is unclear. The lack of an end-to-end baseline is a separate reporting issue, but the physical-consistency premise is more fundamental: a baseline comparison cannot validate the mechanism. I propose an exact rotational-advection test because it isolates the mechanism from numerics and directly compares the OT plan to the true flow map.","tokens_in":20954,"tokens_out":7318,"duration_ms":74366,"concrete_test":"Use a passive scalar advected by a prescribed rotational velocity field (e.g., solid-body rotation of a non-radially-symmetric initial condition) with an exact reference solution. Select two checkpoints spanning a substantial rotation, compute the entropic OT plan (Eq. 13) and the displacement interpolation (Eq. 14) at alpha=0.5, and compare to the exact field. Also compare the OT plan to the true flow map obtained by integrating particle trajectories. If the interpolated field deviates by more than the training time-discretization error from the exact solution, the transport assumption fails and the central claim of physically consistent augmentation collapses.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing claim is physical consistency of the synthetic snapshots produced by Eq. (14). The construction assumes that, between consecutive checkpoints, the full-order dynamics are well approximated by displacement interpolation along a Wasserstein geodesic, i.e., that the quadratic-cost OT plan from Eq. (13) is close to the true Lagrangian flow map. This holds for potential flows but not for the vortical, diffusion-affected dynamics in the benchmarks: the thermal bubble develops a mushroom with roll-up, and the density current develops Kelvin-Helmholtz billows. For such flows, the OT displacement between two snapshots is an optimal rearrangement, not the physical evolution, so the intermediate fields are synthetic artifacts. The signed-field split worsens this: positive and negative parts of theta-prime are normalized and transported with independent plans, so anomalies can cross, which is impossible for a single advecting velocity in the Euler equations. The paper does not validate the transport assumption directly; Figure 4 shows it is grossly wrong for Nc=2, and the method recovers only by making checkpoint intervals short enough that any interpolation is accurate. The subsequent POD residual correction (Eq. 22) can mask but not fix a large, nonsmooth residual, since the GPR then has to learn the entire missing dynamics.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents a reduced-order modeling framework for time-dependent nonlinear advection-dominated PDEs. From a set of Nc checkpoint snapshots, the authors compute entropically regularized optimal transport plans between consecutive checkpoints and use displacement interpolation (Eq. 14) to synthesize Nsynth intermediate snapshots per interval. A virtual-to-real time mapping (piecewise linear L or regression-based M) labels the synthetic snapshots with physical time, and a POD+GPR correction (Eqs. 22-23) is added to reduce residual bias. The method is applied to the rising thermal bubble and the density current benchmarks for mesoscale atmospheric flow, with θ' as the field of interest. Reported results show that mean interpolation and generalization errors decrease as Nc increases and saturate near the training time-discretization error Edisc, that the MinL2 mapping M outperforms the linear mapping L for small Nc, that POD bases from synthetic data give lower projection errors than checkpoint-only bases, and that the POD correction improves density-current errors.","tokens_in":21270,"tokens_out":6851,"duration_ms":59268,"significance":"The framework addresses a real limitation of linear ROMs for advection-dominated problems, and the error decomposition into Edisc, Einterp, and Egen is a useful methodological contribution. The paper also provides a falsifiable prediction: Egen should saturate at Edisc as Nc grows. The use of open-source software (GEA, MultiScaleOT) and the reporting of computational timings and speed-ups are strengths. However, the evidence for the headline claim of 'improved accuracy' is incomplete: the paper does not compare the full pipeline against a standard ROM without OT augmentation, and the physical consistency of the displacement interpolation is validated only indirectly and only for sufficiently short checkpoint intervals. These gaps make the current version a solid proof-of-concept rather than a fully supported general method.","major_comments":[{"comment":"The central claim of improved ROM accuracy is not supported by an end-to-end comparison against a standard non-intrusive ROM without OT augmentation. Figures 8 and 9 quantify the interpolation/generalization error of the OT surrogate itself and the projection error of POD bases built from Ssynth versus Scheck, and Figure 14 shows the effect of the POD correction only for the density current. Missing is a comparison, under identical training and test data, between the complete pipeline in Eq. (23) and a baseline such as POD-R trained on the Nc checkpoints only, or POD-R trained on the original Ntrain snapshots with piecewise-linear time interpolation. Without that baseline, the reported gains could be attributed to the larger number of snapshots in Ssynth rather than to the OT displacement structure.","section":"Section 6, Figures 8-9"},{"comment":"The construction of synthetic snapshots for signed fields is under-specified. The text says that for θ' the positive and negative parts are normalized and transported separately, but it never states how the two independently transported components are recombined into usynth in Eq. (14). Because the two OT plans are independent, positive and negative anomalies can be transported along different paths and overlap in unphysical ways; the paper should specify the recombination and provide a direct check that intermediate fields satisfy basic consistency conditions, such as reproducing the checkpoints at α=0,1 and not creating spurious extrema beyond the entropy-regularization blurring.","section":"Section 4, Steps (2)-(3) and Eq. (14)"},{"comment":"The physical-consistency assumption behind displacement interpolation is only demonstrated in the limit of short checkpoint intervals. The paper itself describes the Nc=2 trajectory in Figure 4 as 'of limited physical relevance,' and Egen in Figures 7-8 is large until Nc is large enough that the interval length is small. This success for large Nc is consistent with any interpolation scheme becoming accurate on short intervals, so it does not specifically validate the Wasserstein-geodesic model. Please quantify the checkpoint interval Δtc at which Egen reaches Edisc and, for at least one interval, compare the OT displacement plan with the physical flow map obtained from the velocity field.","section":"Section 4, Step (3); Figure 4; Remark 4.2"},{"comment":"The POD-based correction (Eqs. 22-23) is a key component of the framework, but results are reported only for the density current. For the thermal bubble, no corrected Egen or Einterp curves are shown, despite the text stating that the correction 'consistently reduces the error.' Please add the corrected-error curves for the thermal bubble, or state explicitly that the correction was not applied in that benchmark.","section":"Section 6.2, Figure 14"}],"minor_comments":[{"comment":"The POD minimization should include the orthogonality constraint U^T U = I; as written, the argmin over all matrices in R^{Nh×Nr} is not the standard POD problem and admits non-unique solutions.","section":"Section 2.1, Eq. (3)"},{"comment":"The formula for usynth uses the L1 norm on the checkpoint fields before normalization, but the OT plan in Eq. (13) is computed between normalized measures; the relation between those normalizations and the mass-interpolation factor (A) should be made explicit, especially for the positive/negative split.","section":"Section 4, Eq. (14)"},{"comment":"The expression Ttest = {j + (0.05 ± Δttest) | j=1,...,NT} is ambiguous; please write the two test-point sets explicitly.","section":"Section 6.1, definition of Ttest"},{"comment":"Figure 2 is referenced as a schematic of the methodology but is not described in the text; adding a short explanation of the data flow would improve readability.","section":"Figure 2"}],"recommendation":"major_revision","confidential_remarks":"To the editor: The manuscript is within scope for an applied mathematics or computational science journal. The missing baseline comparison is, in my view, fixable within a major revision, and the physical-consistency concern is real but is a tunable modeling issue rather than a fatal flaw, because the method explicitly depends on Nc and the paper already reports convergence to Edisc. I would not recommend rejection on novelty grounds; the OT-based augmentation idea is sufficiently distinct from shifted-POD and registration methods. The main risk is that without an end-to-end baseline, the paper overclaims the benefit of OT relative to simply having more snapshots."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a useful, honest engineering paper, not a breakthrough. The pipeline is sensible and the reported error behavior is coherent — Egen drops as Nc grows and saturates at Edisc, and the synthetic POD basis beats the checkpoint-only basis. The main gap is that the paper never runs the natural control experiment: the same POD-R pipeline without the OT augmentation. Without that, the headline \"improved accuracy\" is asserted rather than demonstrated.\n\nWhat is genuinely new: previous displacement-interpolation ROM work (Iollo/Taddei, Cucchiara et al.) was restricted to Gaussian OT maps. Here the entropic OT plan is computed directly on the discrete fields, with positive/negative splitting, checkpoint intervals, a virtual-to-real time mapping (linear and MinL2), and a GPR residual correction. That combination is new, and the modular design means the data augmentation part can be reused by other non-intrusive ROMs.\n\nThe experiments are mostly well done. The two atmospheric benchmarks are appropriate. The observation that Egen saturates at the time discretization error is a good sanity check. The speed-up figures are honest — online inference is cheap, offline OT solves dominate but are parallelizable. The authors also deserve credit for explicitly noting that Nc=2 gives a physically irrelevant trajectory (Figure 4) and for being upfront about it.\n\nThe soft spots are real but not fatal. First, no end-to-end baseline: the POD-basis comparison is not the same as showing the full ROM is more accurate than a POD-R without OT. Second, key implementation details are missing — epsilon for the entropic OT, GPR hyperparameters, and exactly how the signed-field split is done (threshold, handling of zeros). Third, the load-bearing transport assumption is not validated directly. For the thermal bubble and density current, the dynamics include vortical roll-up and diffusion, so the quadratic-cost OT plan is not the physical flow map. The paper recovers by choosing enough checkpoints, which is reasonable, but it means the synthetic snapshots are better described as geometrically plausible intermediates than as physically consistent states. The independent transport of positive and negative anomalies can also allow unphysical crossing, though the error metrics suggest the practical impact is limited. The residual correction presumably hides some of this, but the paper doesn't quantify how much.\n\nWho should read it: people working on non-intrusive ROMs for advection-dominated problems and anyone thinking about data augmentation for limited-data ROM training. It deserves peer review. The referee should ask for the no-OT baseline, full parameter tables, and ideally code/data release.","headline":"A useful, honest OT-based data augmentation ROM with coherent error trends, but the improved-accuracy claim needs a proper no-OT baseline and the transport assumption only holds with dense checkpoints.","tokens_in":21771,"tokens_out":3239,"would_cite":true,"duration_ms":35262,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49Q22","65M08"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that displacement interpolation along optimal-transport geodesics between checkpoint snapshots generates physically consistent synthetic data that, combined with a virtual-to-real time mapping and a Gaussian-process…","keywords":["Reduced-order models","Optimal transport","Displacement interpolation","Wasserstein distance","Data augmentation","Gaussian process regression","Atmospheric flow simulation","Advection-dominated dynamics"],"falsifier":"Run a pair of consecutive checkpoints that straddle a feature which rotates in place or diffuses symmetrically (for example, a stationary diffusing Gaussian bump or a vortex dipole), compute the displacement-interpolated midpoint at $\\alpha=1/2$, and compare it with the full-order solution at the corresponding physical time; the transport-only hypothesis is what fails if the midpoint error remains far above the time-discretization error and the POD/GPR correction cannot close the gap.","tokens_in":20764,"feed_emoji":"🌀","tokens_out":7669,"duration_ms":65263,"temperature":0.7,"pith_summary":"This paper proposes that optimal transport's displacement interpolation—the geodesic path between two probability measures in Wasserstein space—can generate synthetic solution snapshots for training reduced-order models of nonlinear, advection-dominated PDEs. The authors' claim is that interpolating between selected checkpoints along these geodesics yields intermediate states that are physically consistent enough to enrich a POD basis, allowing accurate predictions at arbitrary times even when only coarse and sparse training data are available. They add a virtual-to-real time mapping to attach physical time labels to the synthetic snapshots, and a Gaussian-process regression correction that learns the residual between the interpolated states and the true solution. On two atmospheric benchmarks, a rising thermal bubble and a density current, the method is shown to reduce interpolation and generalization errors and to deliver online speed-ups of roughly two orders of magnitude relative to the full-order solver.","feed_headline":"Wasserstein geodesics sharpen reduced-order models of nonlinear flows","feed_subtitle":"Transport geodesics between snapshots supply synthetic data and finer time scales for advection-dominated flow models.","key_machinery":"The workhorse is displacement interpolation between probability measures, defined in Eq. (14): given an optimal transport plan $\\pi_\\epsilon$ between two checkpoints, the intermediate state at virtual time $\\alpha$ is $\\left((1-\\alpha)\\|u_{k_i}\\|_1 + \\alpha\\|u_{k_{i+1}}\\|_1\\right)\\sum_{l,m}(\\pi_\\epsilon)_{lm}\\delta_{\\mathrm{proj}_X((1-\\alpha)x_l+\\alpha x_m)}$, with positive and negative parts of the field interpolated separately when the field changes sign. The entropic regularization of the transport plan in Eq. (13) makes the plan computable by Sinkhorn-type iterations while remaining close to a true geodesic. Two auxiliary mechanisms carry the online phase: a virtual-to-real time map $F\\in\\{L,M\\}$, where $L$ is piecewise linear and $M$ is a least-squares regression of the global interpolation parameter against physical time, and a POD/GPR residual corrector in Eqs. (21)–(23) that predicts the POD coefficients of the bias $\\mathbf{r}_k = u^k_h - u_{\\mathrm{synth}}(F_1(t_k),F_2(t_k))$.","core_discovery":"The central discovery is that the slow decay of the Kolmogorov n-width that plagues linear ROMs for advection-dominated problems can be partially circumvented at the data level: instead of approximating the solution manifold only with a linear subspace of snapshots, one enriches the snapshot set with displacement interpolants that trace Wasserstein geodesics between checkpoints. For scalar fields the method splits the field into positive and negative parts, normalizes each to unit mass, solves the entropically regularized optimal transport problem between consecutive checkpoints, and pushes each mass particle along a straight line per Eq. (14), with a mass-interpolation factor and a projection back to the computational grid. The paper argues these synthetic snapshots capture genuinely transported features—fronts, rising bubbles, cold-air lobes—so the resulting POD basis is more expressive per mode than one built from checkpoints alone. The residual between geodesic interpolation and the true trajectory is then modeled by POD with Gaussian process regression, which the paper presents as the component that recovers fine-scale structures such as Kelvin–Helmholtz billows.","pith_inferences":["Implicit extension: the data augmentation step is separable from the rest of the pipeline, so any non-intrusive ROM—including neural-network regressors that need large training sets—could consume the synthetic snapshot matrix $S_{\\mathrm{synth}}$ directly.","The transport-only assumption suggests the method should transfer well to Lagrangian or interface-tracking formulations, but it also predicts failure modes: checkpoint intervals containing significant diffusion, chemical reaction, or vortex rotation would need denser checkpoints or an extra non-transport correction.","An adaptive checkpoint selection rule driven by the OT cost between adjacent snapshots, or by a local estimate of the Wasserstein distance traveled, is a natural extension that could reduce the number of checkpoints needed on intervals where the solution barely changes.","If the MinL2 map $M$ is interpreted as a learned time-warping, the same construction could be used for temporal super-resolution of experimental image sequences such as PIV data, where no full-order model exists."],"forward_implications":["As the number of checkpoints $N_c$ grows, the interpolation error decreases monotonically and the generalization error saturates at the time-discretization error of the training data, so the method cannot beat the temporal resolution of the data it is trained on.","The optimal regression map $M$ is more accurate than the linear map $L$ when the solution's rate of change is non-uniform, and the two coincide as $N_c$ increases.","A POD basis built from synthetic OT snapshots has lower projection error than a basis built from checkpoints alone, especially when checkpoints are few and the training time step is coarse.","The POD+GPR residual correction reduces both interpolation and generalization errors in the density current test, with the largest gains when few checkpoints are used, and is what recovers fine-scale Kelvin–Helmholtz structures.","Online inference at arbitrary times costs a matrix-vector multiplication with precomputed plans and achieves speed-ups of order $10^2$ over the reference full-order simulation."],"supporting_citations":[{"why":"Foundational reference for optimal transport; supplies the Wasserstein-space setting in which displacement interpolation is a geodesic.","marker":"[29]"},{"why":"Introduces entropic regularization and the Sinkhorn algorithm that make the transport plans in Eq. (13) computationally tractable.","marker":"[39]"},{"why":"Provides the stabilized multiscale scaling algorithm used to solve the entropy-regularized transport problems without numerical instability.","marker":"[40]"},{"why":"McCann interpolation is the theoretical origin of the displacement interpolation used to define synthetic snapshots in Eq. (14).","marker":"[53]"},{"why":"Earlier convex displacement interpolation for data augmentation that this paper extends from Gaussian models to general scalar fields.","marker":"[38]"},{"why":"Supplies the Gaussian process regression model used for the POD-R residual correction in Eqs. (22)-(23).","marker":"[30]"},{"why":"Defines the rising thermal bubble benchmark and its numerical settings; one of the two test cases validating the method.","marker":"[61]"},{"why":"Defines the density current benchmark, the second and harder test case where the residual correction is shown to be necessary.","marker":"[62]"}],"fun_headline_variants":["Transport geodesics enrich data for reduced-order models","Optimal transport feeds ROMs with synthetic snapshots","Wasserstein interpolation boosts ROM accuracy for nonlinear flows","Geodesic snapshots power reduced-order models of nonlinear dynamics","OT-based augmentation captures fronts in reduced-order models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the physics between two chosen checkpoints is, to first order, pure mass transport: every feature at the later checkpoint can be obtained by moving the features at the earlier one along straight lines, with no significant creation, destruction, diffusion, or rotation in between.","fun_headline_variants_meta":{"raw":{"variants":["Transport geodesics enrich data for reduced-order models","Optimal transport feeds ROMs with synthetic snapshots","Wasserstein interpolation boosts ROM accuracy for nonlinear flows","Geodesic snapshots power reduced-order models of nonlinear dynamics","OT-based augmentation captures fronts in reduced-order models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000191,"raw_usage":{"total_tokens":1353,"prompt_tokens":965,"completion_tokens":388,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":581,"completion_tokens_details":{"reasoning_tokens":312}},"tokens_in":581,"tokens_out":388,"duration_ms":4210,"temperature":1.0,"reasoning_tokens":312,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T21:24:16.175649+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a pair of consecutive checkpoints that straddle a feature which rotates in place or diffuses symmetrically (for example, a stationary diffusing Gaussian bump or a vortex dipole), compute the displacement-interpolated midpoint at $\\alpha=1/2$, and compare it with the full-order solution at the corresponding physical time; the transport-only hypothesis is what fails if the midpoint error remains far above the time-discretization error and the POD/GPR correction cannot close the gap.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Foundational reference for optimal transport; supplies the Wasserstein-space setting in which displacement interpolation is a geodesic."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces entropic regularization and the Sinkhorn algorithm that make the transport plans in Eq. (13) computationally tractable."},{"cited_title":"Cucchiara, A","cited_arxiv_id":null,"evidence_quote":"Earlier convex displacement interpolation for data augmentation that this paper extends from Gaussian models to general scalar fields."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Gaussian process regression model used for the POD-R residual correction in Eqs. (22)-(23)."},{"cited_title":"Ahmad and J","cited_arxiv_id":null,"evidence_quote":"Defines the rising thermal bubble benchmark and its numerical settings; one of the two test cases validating the method."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the density current benchmark, the second and harder test case where the residual correction is shown to be necessary."}],"review_version":1}