{"id":"cda9de78-9d9c-4085-b9e6-83f74938269b","arxiv_id":"2506.09721","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"Generative model latent spaces can reduce geometric parameter dimension and improve the accuracy and speed of reduced-order models for parametrized PDEs.","lead":"This paper proposes using generative models to compress many geometric shape parameters into a few latent variables, then building reduced order models and physics-informed neural networks on those compressed variables. The authors test the idea on Poisson and fluid flow problems over deformed 3D shapes and report better accuracy and faster training.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Out-of-sample leakage in the learned latent parameterization is the load-bearing weakness: latent codes are computed with a generative model trained on every geometry, including the held-out one, so the reported gains may reflect memorization rather than a generalizable reduction of parameter…","rationale":"The reader identified the central weakness as the in-sample fitting of the autoencoder that defines the reduced parameters, and I agree this is the most load-bearing issue. The paper's own error metric in Eqs. (34)-(35) only retrains the ROM, so all reported DROM numbers are conditional on latent codes that have seen the test geometry; the CFFD baseline has no analogous learned representation. This makes the quantitative comparison uninformative about the promised out-of-sample use case. I do not, however, think the paper should be rejected outright: the pipeline is coherent, the PPINN extension is novel relative to GAPINN, and a proper split could validate it. The inconsistency between m=100 and the 600-sample figure captions further supports a request for code/data and a rerun with a clean train/test protocol. Hence the reader's CONDITIONAL verdict stands unchanged.","tokens_in":17213,"tokens_out":6851,"duration_ms":79415,"concrete_test":"Run a proper geometry-level split: hold out a random 20% of geometries before training any generative model; train the autoencoder/generative model only on the remaining 80%, encode the held-out geometries with this fixed model, train each ROM on the training latent codes, and evaluate on the held-out latent codes. Compare with the same split for the CFFD raw parameters, and optionally with a non-generative low-dimensional baseline such as PCA on the point clouds. Repeat over 5-10 seeds/splits and report max/mean error as in Eqs. (34)-(35). If the generative-model advantage over CFFD persists out-of-sample, the central claim is supported; if it shrinks or reverses, the reported gains are an artefact of transductive encoder fitting.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that replacing the 54/81 CFFD parameters with a generative-model latent code improves DROM accuracy and PPINN convergence (Secs. 2.1-2.2). The load-bearing evidence is the leave-one-out error in Eqs. (34)-(35). For the generative-model rows, the test input s_i is computed by an autoencoder trained on all m geometries, including geometry i; only the downstream ROM M_{-i} is retrained. The CFFD baseline has no fitted encoder, so the comparison is asymmetric: the generative pipeline is evaluated on latent codes already optimized to reconstruct the test geometry, whereas the baseline uses raw deformation parameters that were known for every geometry. This does not measure the full encoder+ROM error on a genuinely new geometry, which is exactly the setting promised by 'create new geometries' in the abstract. The same leakage affects the PPINN experiment in Algorithm 2.2: the point-conditioned autoencoder is trained on the same geometries used to define the FEM reference solutions. The exposition also undercuts reproducibility: the DROM text says m=100 for both test cases while the Figs. 2-4 and 6-8 captions say 600 samples, and the first DROM case never states the latent dimension used. The concern is not that the method cannot work; it is that the current experiments do not isolate an out-of-sample benefit of dimensionality reduction.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes using generative models (VAE, AAE, BEGAN, NF, DDPM, EBM) to learn low-dimensional latent encodings of deformed geometries, and then using those latent coordinates as reduced parameters for Data-Driven Reduced Order Models (DROMs) and Parametrised Physics-Informed Neural Networks (PPINNs). The central empirical claim is that replacing the original CFFD/FFD parameters (54 or 81 dimensions) with a generative-model latent code improves the accuracy of DROMs and the convergence of PPINNs, thereby mitigating the curse of dimensionality. Two DROM test cases are reported (Poisson on deformed Stanford Bunnies and two-phase flow around DTCHull), plus one PPINN test case (Laplace on deformed Stanford Bunnies). The paper also contains a self-contained review of divergences and generative-model losses, and an appendix with architectural details.","tokens_in":17488,"tokens_out":4805,"duration_ms":52279,"significance":"If the reported gains were shown to hold for genuinely out-of-sample geometries, the proposal would be practically valuable: a low-dimensional latent parameterization of geometry could make non-intrusive ROMs and PPINNs usable in shape optimization and digital twin settings where the original geometric parameter space is high-dimensional. The paper has concrete strengths: it gives explicit algorithms (Algorithm 2.1 for constrained generation, Algorithm 2.2 for point-conditioned encoding), describes the loss functions of the generative models, and tests on a nontrivial multiphase CFD case. However, the central experimental evidence is currently weakened by an asymmetric evaluation protocol: for the generative-model rows of the leave-one-out errors, the latent code of the held-out sample is produced by a model trained on all samples, whereas the CFFD baseline has no fitted encoder. This prevents the paper from isolating a generalizable out-of-sample benefit of dimensionality reduction. The reported inconsistencies in dataset size and missing latent dimensions further complicate interpretation.","major_comments":[{"comment":"The leave-one-out comparison is asymmetric between the generative-model rows and the CFFD baseline. For a generative-model row, the test input s_i is the latent code produced by an encoder trained on all m geometries, including geometry i, so s_i is an in-sample encoding optimized to reconstruct geometry i; only the ROM M_{-i} is retrained. For CFFD, the test input is the raw deformation parameter that was known for every geometry and is not fitted to the data. Consequently, the reported error reduction does not isolate a generalizable benefit of dimension reduction; it may reflect the autoencoder's ability to reproduce training geometries. To support the abstract's claim about 'create new geometries', the authors need to train the encoder on a split that excludes the test geometry (or use a separate test set of geometries), and evaluate the full encoder+ROM pipeline on those held-out geometries.","section":"§2.1, Eqs. (34)–(35)"},{"comment":"The text states 'In both test cases considered, m=100', but the captions of Figs. 2–4 and 6–8 refer to '600 rabbits' and '600 ships'. The error metrics (34)–(35) depend on m through the max/min normalization and the average, so this is not a purely cosmetic discrepancy. Please clarify which value of m was used for the reported numbers and why the captions differ.","section":"§2.1, after Eq. (35)"},{"comment":"The reduced latent dimension d is never stated for the Stanford Bunny DROM experiment; only the DTCHull case explicitly reports 'The reduced parametrisation (obtained using Generative Models) is 10'. Since d controls the claimed reduction from 54 parameters and directly affects the curse-of-dimensionality argument, this value must be reported for each generative model and each test case.","section":"§2.1, first DROM test case"},{"comment":"The PPINN experiment has the same leakage as the DROM experiment: the point-conditioned autoencoder is trained on the same geometries a_i that are later used to define the FEM reference solutions and the PPINN inputs, and no train/validation split or repeated-seed statistics are reported. This makes it impossible to distinguish the benefit of a lower-dimensional input from the benefit of having fitted the latent representation to the evaluation geometries. Please provide an out-of-sample evaluation and report the mean and standard deviation over multiple seeds.","section":"§2.2, Algorithm 2.2 and Fig. 10"}],"minor_comments":[{"comment":"The manuscript alternates between 'DROMs' and 'DROMS' and between 'PPINNs' and 'PPINNS'; please unify the terminology.","section":"Throughout"},{"comment":"The Adversarial Autoencoder section says 'f_alpha : R^n -> R^k is decoder with weights alpha', but a decoder should map from the latent space R^k to the data space R^n; please correct the domain and codomain.","section":"§1.3.2, Eq. (17) and surrounding text"},{"comment":"The error definitions do not state explicitly what mu_i is for the generative-model rows. Please clarify that mu_i = s_i (the latent code) in those rows and mu_i = gamma_i for the CFFD/FFD baseline.","section":"Eqs. (34)–(35)"},{"comment":"The y-axis tick labels appear garbled in the PDF rendering (for example '3 x 10 1' instead of likely scientific notation); please check the typesetting.","section":"Fig. 10"},{"comment":"In the conclusion, 'Free From Deformation' should be 'Free Form Deformation'.","section":"Section 3"}],"recommendation":"major_revision","confidential_remarks":"The paper's relationship to the authors' prior work [24] should be made explicit; much of the constrained CFFD/generative-model pipeline and the DROM integration appears in [24], while the point-conditioned PPINN extension and the additional test cases are the apparent new contributions. The editor may also wish to ask for the code or data splits to be released, given that the numerical comparison is the central evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe one thing to know about this paper is that its main quantitative claim doesn't hold up as stated. The authors show that reducing 54/81 geometric parameters to a latent code from a generative model improves DROM accuracy and PPINN convergence. But in the leave-one-out ROM evaluation (Eqs. 34-35), the autoencoder that produces the latent code s_i is trained on all m geometries, including the held-out geometry i. Only the downstream ROM M_{-i} is retrained. The CFFD baseline has no fitted encoder, so the comparison pits an encoder that has already seen the test geometry against a raw parameterization that was known for every geometry. That does not measure generalization to a genuinely new geometry, which is what the abstract promises. The PPINN experiment has the same structure: the point-conditioned autoencoder is trained on the same geometries used to generate the FEM reference solutions. So the reported gains may be partly memorization, not a generalizable reduction of parameter dimension.\n\nWhat is new and worth credit: the PPINN variant in Algorithm 2.2. The point-conditioned autoencoder with generative priors over the latent space and coarse-to-fine transfer is a real extension beyond GAPINN, and the comparison with NF/DDPM/EBM is a sensible thing to try. The chapter also gives a compact survey of generative models and a useful description of multilinear constraint enforcement. The authors are honest that the DROM pipeline comes from their CMAME 2024 paper, so the self-citation is not hidden.\n\nThe soft spots beyond the leakage: the paper has internal inconsistencies—the DROM text says m=100 for both test cases while the figure captions say 600 samples, and the latent dimension for the first DROM case is never stated. Section 1 has mathematical sloppiness: Proposition 2's ELBO inequality is written in a way that looks incorrect as stated, the Wasserstein distance is defined with ||·||_2 rather than the usual 1-norm, and the universal approximation claim is phrased imprecisely. These are minor compared to the experimental protocol issue.\n\nWho this is for: researchers working on shape-parameter dimension reduction for non-intrusive ROMs and PPINNs. The PPINN method is worth knowing about, but the experiments need to be redone with a proper train/test split. As it stands, the central claim is plausible but not proven.\n\nMy recommendation: yes, send it to peer review, but require a corrected experimental protocol—train the generative model only on the training geometries, report repeated seeds, resolve the sample count inconsistency, and state the latent dimensions. The paper deserves a serious referee because the PPINN variant is new and the direction is useful, not because the current evidence is conclusive.","headline":"Promising PPINN variant, but the ROM comparison leaks training data into the test latents, so the headline improvement is not cleanly demonstrated.","tokens_in":18047,"tokens_out":3266,"would_cite":false,"duration_ms":31170,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65N99","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"Generative models shrink the parameter space that slows PDE surrogates.","keywords":["generative models","parameter space reduction","reduced order modelling","physics-informed neural networks","geometrically parametrised PDEs","latent space","autoencoders","shape deformation"],"falsifier":"Retrain the autoencoder (or the full generative model) on the training folds only, excluding the held-out geometry from the encoder's training set, and recompute the leave-one-out errors of Eqs. (34)–(35); if the generative parameterisation no longer outperforms the CFFD baseline in that protocol, the reported gain is an evaluation leak rather than genuine dimension reduction.","tokens_in":16969,"feed_emoji":"📐","tokens_out":9939,"duration_ms":100761,"temperature":0.7,"pith_summary":"Solving PDEs on families of deformed 3D shapes is expensive, and surrogate models that map geometric parameters directly to solutions lose accuracy as the number of parameters grows. This paper proposes to replace the raw deformation parameters with the low-dimensional latent codes produced by a generative model, and reports that this substitution improves accuracy and speed for two kinds of surrogates: data-driven reduced order models and parametrised physics-informed neural networks. The evidence includes a Poisson equation on deformed Stanford Bunny meshes, a two-phase naval hull flow with a drag coefficient target, and a Laplace equation solved by PPINNs, with latent dimensions of 5–10 replacing 54–81 original parameters. A sympathetic reader would take the core insight to be that the effective dimensionality of a geometric design family, not the nominal parameter count, is what controls surrogate performance, and that generative models can extract that effective dimensionality from sample geometries.","feed_headline":"Latent codes from generative models sharpen reduced-order PDE solvers","feed_subtitle":"Geometry compression cuts input dimension, improving surrogate accuracy and physics-informed network convergence.","key_machinery":"The load-bearing mechanism is the generative model's latent space used as a reduced parameterisation: each deformed geometry $a_i$ is encoded into $s_i \\in \\mathbb{R}^d$ with $d \\ll l$, where $l$ is the number of free-form-deformation control parameters. For DROMs the encoder is a standard autoencoder-type model (VAE, AAE, or BEGAN) whose decoder output is passed through a multilinear constraint projection, Algorithm 2.1, so that generated shapes keep, for example, a constant barycentre. For PPINNs the mechanism is a point-conditioned autoencoder $g_\\psi : \\mathbb{R}^3 \\times \\mathbb{R}^3 \\to \\mathbb{R}^d$ that maps each point of a reference cloud together with the corresponding point of the deformed cloud to a per-point code, averages these to one $s_i$ per geometry, and lets a decoder $f_\\alpha(s_i, \\cdot)$ rebuild the deformed cloud point by point; a normalising flow, diffusion, or energy-based model then learns the distribution of the $s_i$ so new geometries can be sampled. These codes are the coordinate system in which the map from geometry to PDE solution is learned, and their lower dimension is what the argument credits for the improved accuracy and convergence.","core_discovery":"On its own terms, the paper's central claim is that the latent space of a generative model trained on a family of deformed point clouds is a reduced parameterisation for that family, and that using it in place of the original geometric parameters improves both DROMs and PPINNs. In the DROM experiments, 54-parameter and 81-parameter CFFD deformations are compressed to latent codes, including a 10-dimensional code for the naval hull, and Gaussian process, radial-basis, and tree regressors trained on the codes show lower max and mean errors and lower CPU times than the same regressors trained on the original parameters. In the PPINN experiment, a point-conditioned autoencoder pools per-point features into a 5-dimensional code per geometry, a normalising flow, diffusion model, or energy-based model learns the distribution of those codes, and PPINNs trained on the codes converge to lower loss and lower L1/L2 error against finite-element solutions than PPINNs trained with the 54-parameter FFD parameterisation. The paper attributes the improvement to a softer curse of dimensionality, citing the Gaussian process generalisation-error scaling $m^{-c/\\dim(\\Gamma)}$, and notes that generation under multilinear constraints such as a fixed barycentre remains compatible with the whole pipeline.","pith_inferences":["A stricter evaluation would retrain the autoencoder on each training fold and only then encode the held-out geometry; such a protocol would separate true out-of-sample dimension reduction from the current leave-one-out scheme, in which the encoder has already seen every sample.","The same latent-space reduction should extend from regression surrogates to optimisation and uncertainty quantification over shapes, where gradients or sampling over a 5–10 dimensional latent space are far cheaper than over the original 54–81 dimensional parameter space.","If the curse-of-dimensionality explanation is right, the accuracy gap between latent and original parameterisations should grow with the original parameter count $l$ and narrow as the latent dimension $d$ approaches $l$; that predicted scaling is directly testable.","The point-conditioned encoder is not tied to a fixed mesh, so the approach may generalise to non-matching geometries, adaptive meshes, and time-dependent or coupled GPPDEs where the map from shape to solution is even more expensive."],"forward_implications":["A geometry-to-solution pipeline can be prefixed by a generative compression step, replacing dozens of deformation parameters with a latent code of about one fifth to one tenth of the original dimension.","The same reduced parameterisation lets PPINNs be trained with fewer parameter and point samples, yielding lower loss and better agreement with finite-element solutions.","Because the point-conditioned decoder rebuilds point clouds pointwise, a PPINN solution trained on a coarse mesh can be transferred to a finer mesh without retraining.","Constrained shape generation survives the reduction: multilinear constraints such as a fixed barycentre are enforced on the decoder output, so the latent parameterisation still lives inside the admissible geometry family.","Sampling time for the generative approach is lower than for the full parameterisation; the paper reports about a twofold speed-up for PPINN sampling."],"supporting_citations":[{"why":"Supplies the constrained free-form-deformation sampling, the multilinear-constraint projection algorithm, and the earlier generative model parameter reduction that this chapter extends.","marker":"[24]"},{"why":"Variational autoencoder whose latent codes are used as the reduced parameterisation in the DROM experiments.","marker":"[30]"},{"why":"Adversarial autoencoder, one of the generative models whose latent parameterisation is compared against CFFD.","marker":"[33]"},{"why":"Boundary equilibrium GAN, the third generative model tested in the DROM pipeline.","marker":"[40]"},{"why":"Denoising diffusion probabilistic model used to learn the latent-code distribution in the PPINN pipeline.","marker":"[31]"},{"why":"Normalising flow used to learn and sample the latent distribution for PPINN geometries.","marker":"[41]"},{"why":"Energy-based model used to learn and sample the latent distribution for PPINN geometries.","marker":"[42]"},{"why":"Point-wise feature conditioning that motivates the encoder pooling each point against a reference cloud, enabling variable-topology deformation in the PPINN method.","marker":"[50]"},{"why":"Gaussian process regression baseline whose accuracy, measured leave-one-out, is the DROM comparison showing the benefit of the latent parameterisation.","marker":"[44]"}],"fun_headline_variants":["Generative models compress geometry, boosting reduced-order PDE surrogates","Latent codes shrink parameter space, sharpen PDE solvers","Generative latent spaces reduce PDE parameters, improve surrogates","Parameter compression via generative models enhances reduced-order modeling","Latent geometry codes cut PDE parameter dimension, lift accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's errors are computed leave-one-out for the surrogate model, but the generative model that defines each sample's latent code is trained on all samples, including the held-out one, so the comparison does not isolate out-of-sample dimension reduction.","fun_headline_variants_meta":{"raw":{"variants":["Generative models compress geometry, boosting reduced-order PDE surrogates","Latent codes shrink parameter space, sharpen PDE solvers","Generative latent spaces reduce PDE parameters, improve surrogates","Parameter compression via generative models enhances reduced-order modeling","Latent geometry codes cut PDE parameter dimension, lift accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000354,"raw_usage":{"total_tokens":1955,"prompt_tokens":1002,"completion_tokens":953,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":618,"completion_tokens_details":{"reasoning_tokens":871}},"tokens_in":618,"tokens_out":953,"duration_ms":7716,"temperature":1.0,"reasoning_tokens":871,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:42:27.267193+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the autoencoder (or the full generative model) on the training folds only, excluding the held-out geometry from the encoder's training set, and recompute the leave-one-out errors of Eqs. (34)–(35); if the generative parameterisation no longer outperforms the CFFD baseline in that protocol, the reported gain is an evaluation leak rather than genuine dimension reduction.","supporting_citations":[{"cited_title":"Computer Methods in Applied Mechanics and Engineering, 423, 116823 (2024)","cited_arxiv_id":null,"evidence_quote":"Supplies the constrained free-form-deformation sampling, the multilinear-constraint projection algorithm, and the earlier generative model parameter reduction that this chapter extends."},{"cited_title":"International Conference on Learning Representations (2014)","cited_arxiv_id":null,"evidence_quote":"Variational autoencoder whose latent codes are used as the reduced parameterisation in the DROM experiments."},{"cited_title":"International Conference on Learning Representations (2016)","cited_arxiv_id":null,"evidence_quote":"Adversarial autoencoder, one of the generative models whose latent parameterisation is compared against CFFD."},{"cited_title":"Advances in Neural Information Processing Systems 33 (2020)","cited_arxiv_id":null,"evidence_quote":"Denoising diffusion probabilistic model used to learn the latent-code distribution in the PPINN pipeline."},{"cited_title":"Proceedings of the 32nd International Conference on Machine Learning (2015)","cited_arxiv_id":null,"evidence_quote":"Normalising flow used to learn and sample the latent distribution for PPINN geometries."},{"cited_title":"Ad- vances in Neural Information Processing Systems 32 (2019)","cited_arxiv_id":null,"evidence_quote":"Energy-based model used to learn and sample the latent distribution for PPINN geometries."},{"cited_title":"Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2017)","cited_arxiv_id":null,"evidence_quote":"Point-wise feature conditioning that motivates the encoder pooling each point against a reference cloud, enabling variable-topology deformation in the PPINN method."},{"cited_title":"MIT Press, Cambridge (2006) Generative Models for Parameter Space Reduction applied to Reduced Order Modelling 25","cited_arxiv_id":null,"evidence_quote":"Gaussian process regression baseline whose accuracy, measured leave-one-out, is the DROM comparison showing the benefit of the latent parameterisation."}],"review_version":1}