Pith. sign in

REVIEW 4 major objections 6 minor 44 references

A Generative Model-Free Form Deformation Approach for the Generation of Mesh Motions with Applications to PDE

T0 review · 4 major / 6 minor · reviewed 2026-08-02 · deepseek-v4-flash

Pith's one-line read The paper establishes that any smooth genus-0 surface deformation can be approximated arbitrarily well by the time-one map of an ODE with free-form-deformation drift, and that the resulting latent representation speeds up reduced-order PDE

desk verdict Solid theory, overstated numerics: the ODE–FFD deformation framework is a genuine new tool, but the generative model's benefits are not isolated from plain dimensionality reduction. read the letter →

arxiv 2607.13202 v1 pith:KIFH34LX submitted 2026-07-14 math.NA cs.NA

classification math.NAcs.NA MSC 65D1765M6068T05
keywords free-formdeformationordinarydifferentialequationsshapegenus-0surfacesuniversalapproximationproperorthogonaldecompositionnormalizingflowsreduced-ordermodels
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper introduces FFD-ODE, modeling a deforming mesh as trajectories of an ODE whose velocity field is a time-dependent Bernstein-polynomial (Free Form Deformation) field driven by control-lattice displacements. It proves a universality result: for any closed smooth genus-0 surfaces inside the unit cube and any tolerance, there exist Bernstein degrees and boundary-anchored control functions making the Chamfer distance between the time-one image of the source and the target arbitrarily small. It further proves that the map from control trajectories to flows on the cube is a bijection with two-sided quantitative Lipschitz bounds, so compressing controls via Proper Orthogonal Decomposition cannot create spurious deformation fields. On two CFD benchmarks (a deforming bunny in incompressible flow and a deforming sphere in hypersonic flow), reduced-order models built on the learned latent trajectories cut surrogate errors by 30–40% and accelerate a drag-minimization loop by about three orders of magnitude.

What carries the argument

The central object is the FFD-ODE system ẋ = F_{α(t)}(x), where F is the trivariate Bernstein polynomial interpolation of the time-dependent control displacements α(t), with boundary control components fixed to zero so the cube's boundary is invariant. Three tools carry the argument: Bernstein polynomial universality to approximate the diffeotopy's velocity field, Grönwall's inequality to bound the resulting flow error, and a coercivity estimate for the pulled-back Gram matrix M_α(t) = ∫ B(y_α(t,x)) B(y_α(t,x))ᵀ dx, whose minimal eigenvalue is bounded below explicitly; this last estimate is what gives the reverse Lipschitz inequality and the bijection.

What would settle it

For m=n=o=2 and a smooth control α, compute the smallest eigenvalue of the pulled-back Gram matrix M_α(t) at t=1/2 and compare it with the constant exp(-2√3 ¯m ∥α∥_L1) / [(2m+1)(2m choose m)(2n+1)(2n choose n)(2o+1)(2o choose o)]; if the eigenvalue falls below this bound, Theorem 4's reverse inequality is violated and the claimed bijection would need revisiting.

Watch

Extended reading notes

Core claim

Deformations between genus-0 meshes of different connectivity can be encoded as the flow of an ODE whose drift is a time-dependent Free Form Deformation field; this encoding is universal in the Chamfer metric, and the map from control trajectories to flows is a bijection with quantitative Lipschitz equivalences. The universality proof extends the surface isotopy to a compactly supported diffeotopy of the cube, approximates the diffeotopy's velocity field by Bernstein polynomials, and applies Grönwall's inequality to transfer the field error to the flow. The bijection proof relies on a coercivity estimate for the Gram matrix of the Bernstein basis pulled back by the flow; the minimal eigenval

Load-bearing premise

The universality proof assumes the Bernstein approximant of the diffeotopy's velocity field, with coefficients sampled at the control lattice, has exactly zero boundary control components while retaining sup-norm accuracy; this holds if the field vanishes on the boundary, but a fully detailed proof of the coefficient constraint is not given, and the numerical results depend on the unverified assumption that fixed-boundary 5×5×5 control lattices produce stable, non-degenerate

Editorial extensions

If this is right

  • Any smooth deformation of a genus-0 shape can be represented, to arbitrary Chamfer accuracy, by a low-dimensional ODE control, making mesh morphing independent of matching mesh graphs.
  • Because the control-to-flow map is a bijection with two-sided Lipschitz bounds, dimension reduction (POD) in control space is faithful and cannot generate spurious deformation fields.
  • Reduced-order models built on the generative latent space achieve 30–40% lower leave-one-out errors on quantity-of-interest prediction than ROMs on raw FFD-ODE controls, in both incompressible and hypersonic benchmarks.
  • The FFD-ODE formulation gives injective, topology-preserving deformations by construction, which keeps moving meshes valid during PDE solution.
  • The generative prior regularizes the inverse deformation problem: an unconstrained 10-dimensional drag-minimization converges in about 5 seconds, while a constrained raw-control optimization stalls at 1000 iterations and takes about 3 hours.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same diffeotopy-extension argument should extend to any pair of diffeomorphic surfaces, not only genus-0, whenever the ambient isotopy is supported inside the cube; the genus-0 restriction appears to come from the chosen isotopy theorem, not from the ODE-FFD mechanism itself.
  • The explicit coercivity constant in the reverse Lipschitz bound suggests a computable criterion for how many POD modes are needed to guarantee a given deformation accuracy, turning the qualitative bijection into a practical mode-selection rule.
  • Because the velocity field is linear in the controls, the forward map from α to the flow is well-suited to differentiable programming and adjoint-based optimal control, which could accelerate the initial trajectory-fitting stage beyond the forward–backward sweep used here.
  • The generative prior's strong regularization effect on drag optimization hints that the same latent representation could regularize other ill-posed inverse shape problems, such as aerodynamic shape optimization or medical image registration, where mesh validity is the bottleneck.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper introduces FFD-ODE, a framework for deforming genus-0 surfaces by modelling the deformation as the time-1 map of an ODE whose drift is a time-dependent Free Form Deformation (FFD). The main theoretical contribution is Theorem 2, which claims universal approximation of smooth genus-0 surface deformations in Chamfer distance using FFD-ODE flows with boundary-anchored control coefficients. The proof combines Bernstein polynomial universality, Hirsch's diffeotopy extension, and Grönwall's inequality. The paper also proves quantitative Lipschitz-type equivalences between control trajectories and their flows (Theorems 3–4), which are used to justify POD reduction of the control space. A generative model based on POD and STARFlow is then introduced to produce a low-dimensional latent parametrisation of deformation trajectories. The method is tested on two CFD benchmarks: incompressible flow past a deforming Stanford Bunny (FEniCSx) and hypersonic flow past a deforming sphere (OpenFOAM). Reduced-order models (KNN, GPR, RF) are trained either on the original high-dimensional FFD-ODE control trajectories or on the STARFlow latent variables, and their leave-one-out cross-validation errors are compared.

Significance. If the claims hold, the paper makes a useful theoretical contribution: it rigorously connects classical FFD-based geometric morphing with ODE flows, proving an expressivity guarantee for genus-0 surfaces and providing quantitative bounds that justify model-order reduction in the control space. The bi-Lipschitz bounds of Theorems 3–4 are valuable and appear essentially correct. The numerical section, however, does not establish the paper's central practical claim that the generative model improves ROM accuracy: the comparison conflates dimensionality reduction with the specific STARFlow generative mechanism. The reported accuracy gains and speed-ups are also overstated relative to the tables and the optimisation setup. These issues are fixable with additional baseline experiments and more careful reporting, so the paper warrants a major revision rather than rejection.

major comments (4)
  1. [Section 5, Tables 1–4] The central numerical claim — that ROMs trained on STARFlow latent variables are more accurate than ROMs trained on FFD-ODE control trajectories — is confounded by dimensionality. The FFD-ODE ROMs receive the full 375-dimensional control trajectories, while the GEN ROMs receive 10-dimensional STARFlow latents. Any dimensionality reduction, including plain linear POD, is expected to improve LOOCV accuracy of non-intrusive ROMs on a small dataset. To attribute the improvement to the generative model, the authors must add a baseline consisting of ROMs trained on the first D=10 POD coefficients of the same FFD-ODE trajectories, without the autoencoder or STARFlow. Without this baseline, the statement in Section 5 that the lower errors reflect 'the reduced parametrisation afforded by the generative model' is unsupported.
  2. [Section 6 (Conclusion)] The conclusion claims a '30–40% reduction in error' relative to classical high-dimensional parametrisations. The largest reduction in Tables 1–4 is about 28% (Table 2, RF L1: 0.1354 to 0.0979, a 27.7% reduction). Several entries are far smaller (e.g., Table 3, GPR L2: 0.1620 to 0.1540, about 4.9%). The stated range is not supported by the reported numbers and should be corrected.
  3. [Section 5.2 and Figure 14] The claimed '3600× speed-up' is not a fair comparison. The STARFlow-based optimisation is reported to converge in 394 iterations and about 5 seconds, while the FFD-ODE-based optimisation is truncated at its maximum of 1000 iterations without convergence after about 3 hours. These are not equivalent termination conditions. Moreover, 3 hours divided by 5 seconds is approximately 2160×, not 3600×. The authors should report time-to-convergence under the same stopping criterion, or at least report per-iteration costs and clearly state that the baseline did not converge.
  4. [Section 3.1, after Theorem 4] The text states: 'This result implies that there is a bijection between a control-points trajectory α∈V and its corresponding solution y∈W.' Theorem 4 provides injectivity and a quantitative lower bound, but it does not establish surjectivity onto the full space W. If the intended meaning is a bijection between V and the image {y_α : α∈V}, this is tautological and should be stated as such; if surjectivity onto W is claimed, it requires proof. The injectivity part is sufficient for the POD argument and should be phrased accordingly.
minor comments (6)
  1. [Appendix A, proof of Theorem 2] The final line asserts that because g is supported in (0,1)^3, the Bernstein coefficients at boundary lattice points are zero. This is correct if 'supported' is understood as compact support away from the boundary, but the proof should state that the lattice resolution is chosen so that every boundary knot lies outside the support of g. Without that explicit statement, the step reads as an assertion rather than a derivation.
  2. [Section 1.2] Typo: 'The main differences are the differences are that' should be 'The main differences are that'.
  3. [Section 3.1] Typo: 'Our objective is to demonstrate that is to show that' should be 'Our objective is to show that'.
  4. [Section 4.1, governing equations] The initial condition is written as u(0,x)=0 only on ∂Ω(0), which is insufficient for the incompressible Navier–Stokes equations. It should likely be u(0,x)=0 for x∈Ω(0) or a suitable divergence-free initial field; please clarify.
  5. [Section 4] The text says 'the number of control points is set to 4 along each spatial axis' but later refers to '5×5×5 velocity fields of control points'. Clarify whether 4 is the polynomial degree (giving 5 control points per axis) or the number of control points. The notation m,n,o in Section 2 suggests degree, but the wording is ambiguous.
  6. [Tables 1–4] No measure of variability (e.g., repeated training with different splits, or error bars) is reported. With 100 samples and default Scikit-Learn hyperparameters, some of the differences (especially in Table 3) may not be statistically meaningful; a paired test or repeated CV would strengthen the comparison.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the universal-approximation and bijection results are self-contained derivations from external classical results; the generative-model comparison is a statistical evaluation rather than a definitional reduction.

full rationale

The derivation chain is not circular. Theorem 2 is proved from classical Bernstein approximation (Theorem 1, ref. [37]) plus Hirsch's ambient-isotopy extension theorem (ref. [43]); the boundary-coefficient condition α_ijk=0 follows because the Bernstein coefficients evaluate g at boundary knots, where g vanishes due to the support of the diffeotopy in (0,1)^3. This is a mathematical consequence, not an imported assumption. Theorems 3 and 4 are proved from elementary Bernstein-basis inequalities, Proposition 1, and the external Lyche-Scherer eigenvalue bound (ref. [44]); they do not assume the POD conclusion they are used to justify. The POD and STARFlow sections apply standard dimension reduction and an external normalizing-flow architecture (refs. [12, 13]). The authors' own prior works (refs. [5, 6]) appear only as motivating citations and do not carry the load of any theorem. The numerical ROM comparisons and QoI-distribution checks evaluate a fitted generative model against the distribution it was trained on; the reported accuracy gains would be strengthened by a POD-only baseline, and the '30-40%' and '3600x' statements somewhat overstate the tabulated results, but these are statistical/comparison concerns, not circular reductions. No equation in the paper is equivalent to its own input by construction.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

No new physical entities are introduced; the framework's free parameters are all methodological (regularization weight, POD truncation, latent dimension, control-grid resolution, constraint thresholds). The load-bearing external inputs are classical theorems (Hirsch, Bernstein, Lyche–Scherer) plus one unproved boundary-coefficient step in the universality proof. The numerical claims additionally depend on unstated hyperparameters of the autoencoder, STARFlow, and the PDE solvers.

free parameters (5)
  • Regularization parameter ρ in the control-trajectory optimization
    The Chamfer-matching objective (Section 2) includes ρ||α(t)−β_0+c||²; ρ is introduced without specification, and its value affects the resulting trajectories and hence all downstream ROM results.
  • POD retained modes M = 80 = 80
    Section 4.1 states 'we retain the first M=80 POD modes' — a hand-chosen truncation that directly sets the reconstruction quality of control trajectories.
  • Autoencoder latent dimension D = 10 = 10
    Section 3.2 fixes latent dimension D=10; all 'GEN' results depend on this choice, and no sensitivity study is reported.
  • FFD control lattice resolution (4 along each spatial axis / 5×5×5 grids) = 5x5x5
    Section 4 sets the control grid to 4 (then 5×5×5) per axis; the expressivity and the mesh-quality behavior are sensitive to this resolution, and no resolution study is given.
  • Optimization constraint thresholds (||α(i)−α(i−1)|| ≤ 0.01, RMS bound 1.65) = 0.01 / 1.65
    The FFD-ODE drag-minimization constraints in Section 4.2 are hand-chosen bounds to 'prevent non-physical deformations'; they change the feasible set and the comparison with the unconstrained GEN run.
assumptions (6)
  • standard math Hirsch's isotopy-extension theorem: any two closed genus-0 surfaces in (0,1)^3 admit a diffeotopy of R^3 supported in (0,1)^3 connecting them.
    Invoked as [38] and [43], Theorem 1.3, p.180 in the proof of Theorem 2 (Appendix A). This provides the vector field g whose approximation yields the universality claim.
  • standard math Classical Bernstein uniform approximation of continuous functions on the cube by FFD maps (Theorem 1, cited to [37]).
    Used in Appendix A to approximate the diffeotopy vector field g by a Bernstein-FFD field with controlled sup error.
  • standard math Minimum eigenvalue of the univariate Bernstein mass matrix is 1/[(2m+1) C(2m,m)].
    Invoked as [44] in Proposition 4 to obtain the coercivity constant ω_γ for the deformed Gram matrix, which underpins the lower bound in Theorem 4.
  • ad hoc to paper Unproven step in Theorem 2's proof: the Bernstein-FFD approximation of the supported vector field g can be chosen with exactly zero coefficients at boundary lattice indices while preserving sup-norm accuracy.
    Appendix A states 'since g is supported in (0,1)^3, the coefficients associated with boundary grid points are necessarily zero' without a proof. This is plausible because g vanishes in a neighborhood of ∂[0,1]^3, but it is the structural premise for the boundary-anchoring of the flow (Proposition 2) and for the stated zero-boundary-coefficient constraint in Theorem 2.
  • domain assumption In the fluid benchmarks, the FFD-ODE velocity field restricted to the box boundary is identically zero, so the deforming domain remains inside the box.
    Section 4.1: 'the domain remains in the box as F_α restricted to the box boundary is 0.' This is only guaranteed if boundary control coefficients vanish, which is imposed (not proved) in the numerical setup.
  • domain assumption The flow of a time-dependent polynomial vector field is a C¹ diffeomorphism of the cube, giving guaranteed injectivity of the mesh deformation.
    Used throughout (Propositions 2–3) and advertised as the key advantage over prior morphing methods; requires the boundary-vanishing property and boundedness of the control trajectory.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Generative Model-Free Form Deformation Approach for the Generation of Mesh Motions with Applications to PDE." pith.science (2026). https://pith.science/paper/KIFH34LX

@misc{pith2026260713202,
  author       = {Pith},
  title        = {Pith review of: A Generative Model-Free Form Deformation Approach for the Generation of Mesh Motions with Applications to PDE},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KIFH34LX}},
  note         = {Machine review of arXiv:2607.13202}
}
read the original abstract

We introduce a topology-agnostic framework for matching deformations of three-dimensional shapes with non-isomorphic mesh graphs by modelling the deformation as the flow of an Ordinary Differential Equation (ODE). The velocity field is parameterised by a time-dependent Free Form Deformation (FFD), expressed through displacements of a coarse control lattice, yielding a smooth and low-dimensional representation that decouples the deformation model from the discretisation of the source and target surfaces. Under mild regularity assumptions, we prove that the induced ODE map is a universal approximator (in the sup norm) for mappings between genus-0 surfaces, providing a theoretical expressivity guarantee. To further compress the representation and enable probabilistic inference, we couple the ODE--FFD model with a flow-based generative approach in the TarFlow framework, learning a compact latent parametrisation over time series of FFD maps. The resulting method supports efficient sampling and optimisation of plausible deformation trajectories while preserving mesh quality, and it enables scalable reduced-order modelling. Experiments on deforming-body flow benchmarks demonstrate improved accuracy and computational efficiency of reduced-order models constructed from the learned latent dynamics.

Figures

Figures reproduced from arXiv: 2607.13202 by the authors.

Figure 1
Figure 1. Left: source bunny model exhibiting the topology to be transferred. [PITH_FULL_IMAGE:figures/full_fig_p010_1.png] view at source ↗
Figure 2
Figure 2. Deformation of the bunny in time. 11 [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Position of the control points in time. 12 [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: Summary of the training pipeline. 16 [PITH_FULL_IMAGE:figures/full_fig_p016_4.png]
Figure 5
Figure 5. Figure 5: Workflow for the construction of the ROM by generating snapshots [PITH_FULL_IMAGE:figures/full_fig_p018_5.png]
Figure 6
Figure 6. Figure 6: Section of the computational mesh used for the Stanford Bunny test [PITH_FULL_IMAGE:figures/full_fig_p021_6.png]
Figure 7
Figure 7. Figure 7: Detail of the computational mesh adopted for the sphere test case. [PITH_FULL_IMAGE:figures/full_fig_p028_7.png]
Figure 8
Figure 8. Figure 8: First row: bunny deformation at time t = 1 obtained with FFD￾ODE (first two panels) and with the generative model (last panel). Second row: corresponding Navier–Stokes velocity fields. Marked discrepancies among the solutions can be observed. 29 [PITH_FULL_IMAGE:figur…
Figure 9
Figure 9. Figure 9: Distribution of the average velocity norm for the FFD-ODE outputs [PITH_FULL_IMAGE:figures/full_fig_p030_9.png]
Figure 10
Figure 10. Figure 10: Distribution of the root mean square control-point deformation for the [PITH_FULL_IMAGE:figures/full_fig_p031_10.png]
Figure 11
Figure 11. Figure 11: First row: temperature field at time t = 1 from FFD-ODE (first two panels) and from the generative model (last panel). Pronounced variability in the solutions can be observed. 31 [PITH_FULL_IMAGE:figures/full_fig_p031_11.png]
Figure 12
Figure 12. Figure 12: Distribution of drag for the FFD-ODE outputs and the STARFlow [PITH_FULL_IMAGE:figures/full_fig_p032_12.png]
Figure 13
Figure 13. Figure 13: Distribution of the root mean square control-point deformation for the [PITH_FULL_IMAGE:figures/full_fig_p033_13.png]
Figure 14
Figure 14. Figure 14: Optimisation performance for drag minimisation with FFD-ODE [PITH_FULL_IMAGE:figures/full_fig_p034_14.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 3 linked inside Pith

  1. [1]

    Hull Shape Design Optimization with Parameter Space and Model Reductions, and Self-Learning Mesh Morphing

    Nicola Demo, Marco Tezzele, Andrea Mola, and Gianluigi Rozza. Hull Shape Design Optimization with Parameter Space and Model Reductions, and Self-Learning Mesh Morphing. Journal of Marine Science and Engineering, 9(2):185, February 2021

  2. [2]

    Marsden, and Simone Deparis

    Riccardo Tenderini, Luca Pegolotti, Fanwei Kong, Stefano Pagani, Francesco Regazzoni, Alison L. Marsden, and Simone Deparis. Deformable registration and generative modelling of aortic anatomies by auto-decoders and neural ODEs. npj Biological Physics and Mechanics, 2(1):26, December 2025

  3. [3]

    A hybrid reduced-order model for segregated fluid-structure interaction solvers in an ALE approach at high Reynolds number

    Valentin Nkana Ngan, Giovanni Stabile, Andrea Mola, and Gianluigi Rozza. A hybrid reduced-order model for segregated fluid-structure interaction solvers in an ALE approach at high Reynolds number. Computers & Mathematics with Applications, 180:299–321, February 2025

  4. [4]

    Sederberg and Scott R

    Thomas W. Sederberg and Scott R. Parry. Free-form deformation of solid geometric models. ACM SIGGRAPH Computer Graphics, 20(4):151–160, 1986

  5. [5]

    Generative models for the deformation of industrial shapes with linear geo- metric constraints: Model order and parameter space reductions

    Guglielmo Padula, Francesco Romor, Giovanni Stabile, and Gianluigi Rozza. Generative models for the deformation of industrial shapes with linear geo- metric constraints: Model order and parameter space reductions. Computer Methods in Applied Mechanics and Engineering, 423:116823, 2024

  6. [6]

    Generative Models for Param- eter Space Reduction applied to Reduced Order Modelling, June 2025

    Guglielmo Padula and Gianluigi Rozza. Generative Models for Param- eter Space Reduction applied to Reduced Order Modelling, June 2025. arXiv:2506.09721

  7. [7]

    Salmoiraghi, A

    F. Salmoiraghi, A. Scardigli, H. Telib, and G. Rozza. Free-form deforma- tion, mesh morphing and reduced-order methods: enablers for efficient aero- dynamic shape optimisation. International Journal of Computational Fluid Dynamics, 32(4-5):233–247, May 2018

  8. [8]

    Fluid-structure Interactions, volume 118

    Thomas Richter. Fluid-structure Interactions, volume 118. Springer Inter- national Publishing, 2017

Show all 44 references
  1. [9]

    Exact and Approximate Controllability for Distributed Parameter Systems

    Roland Glowinski, Jacques-Louis Lions, and Jiwen He. Exact and Approximate Controllability for Distributed Parameter Systems. Cambridge University Press, March 2008. 35

  2. [10]

    G. Cybenko. Approximation by superpositions of a sigmoidal function. Mathematics of Control, Signals, and Systems, 2(4):303–314, December 1989

  3. [11]

    Approximation capabilities of multilayer feedforward networks

    Kurt Hornik. Approximation capabilities of multilayer feedforward networks. Neural Networks, 4(2):251–257, 1991

  4. [12]

    STARFlow: Scaling Latent Normalizing Flows for High- resolution Image Synthesis, June 2025

    Jiatao Gu, Tianrong Chen, David Berthelot, Huangjie Zheng, Yuyang Wang, Ruixiang Zhang, Laurent Dinh, Miguel Angel Bautista, Josh Susskind, and Shuangfei Zhai. STARFlow: Scaling Latent Normalizing Flows for High- resolution Image Synthesis, June 2025. arXiv:2506.06276

  5. [13]

    Normalizing Flows are Capable Generative Mod- els, June 2025

    Shuangfei Zhai, Ruixiang Zhang, Preetum Nakkiran, David Berthelot, Ji- atao Gu, Huangjie Zheng, Tianrong Chen, Miguel Angel Bautista, Navdeep Jaitly, and Josh Susskind. Normalizing Flows are Capable Generative Mod- els, June 2025. arXiv:2412.06329

  6. [14]

    Brooks and Thomas J

    Alexander N. Brooks and Thomas J. R. Hughes. Streamline upwind/Petrov- Galerkin formulations for convection dominated flows with particular em- phasis on the incompressible Navier-Stokes equations. Computer Methods in Applied Mechanics and Engineering, 32(1-3):199–259, September 1982

  7. [15]

    Baratta, Joseph P

    Igor A. Baratta, Joseph P. Dean, Jørgen S. Dokken, Michal Habera, Jack S. Hale, Chris N. Richardson, Marie E. Rognes, Matthew W. Scroggs, Nathan ORCID icon Sime, and Garth N. Wells. DOLFINx: The next generation FEniCS problem solving environment. DOLFINx: The next generation F...

  8. [16]

    Scroggs, Igor A

    Matthew W. Scroggs, Igor A. Baratta, Chris N. Richardson, and Garth N. Wells. Basix: a runtime finite element basis evaluation library. Journal of Open Source Software, 7(73):3982, May 2022

  9. [17]

    Alnæs, Anders Logg, Kristian B

    Martin S. Alnæs, Anders Logg, Kristian B. Ølgaard, Marie E. Rognes, and Garth N. Wells. Unified form language. ACM Transactions on Mathematical Software, 40(2):1–37, February 2014

  10. [18]

    Zippered polygon meshes from range images

    Greg Turk and Marc Levoy. Zippered polygon meshes from range images. In Proceedings of the 21st annual conference on Computer graphics and interactive techniques - SIGGRAPH ’94, pages 311–318, Not Known, 1994. ACM Press. 36

  11. [19]

    H. G. Weller, G. Tabor, H. Jasak, and C. Fureby. A tensorial approach to computational continuum mechanics using object-oriented techniques. Computers in Physics, 12(6):620–631, November 1998

  12. [20]

    Anderson

    John D. Anderson. Hypersonic and High-Temperature Gas Dynamics, Third Edition. American Institute of Aeronautics and Astronautics, Inc., March 2019

  13. [21]

    Reduced Basis Methods for Partial Differential Equations, volume 92

    Alfio Quarteroni, Andrea Manzoni, and Federico Negri. Reduced Basis Methods for Partial Differential Equations, volume 92. Springer Interna- tional Publishing, 2016

  14. [22]

    Willcox and J

    K. Willcox and J. Peraire. Balanced Model Reduction via the Proper Or- thogonal Decomposition. AIAA Journal, 40(11):2323–2330, November 2002

  15. [23]

    Mesh Deformation Approaches – A Survey

    Selim Mm and Koomullil Rp. Mesh Deformation Approaches – A Survey. Journal of Physical Mathematics, 7(2), 2016

  16. [24]

    Auto-Encoding Variational Bayes, 2013

    Diederik P Kingma and Max Welling. Auto-Encoding Variational Bayes, 2013

  17. [25]

    Variational Inference with Normaliz- ing Flows

    Danilo Rezende and Shakir Mohamed. Variational Inference with Normaliz- ing Flows. In Proceedings of the 32nd International Conference on Machine Learning, pages 1530–1538. PMLR, June 2015

  18. [26]

    Generative ad- versarial networks

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde- Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative ad- versarial networks. Commun. ACM, 63(11):139–144, October 2020

  19. [27]

    Denoising Diffusion Proba- bilistic Models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Proba- bilistic Models. In Advances in Neural Information Processing Systems, volume 33, pages 6840–6851. Curran Associates, Inc., 2020

  20. [28]

    ShapeFusion: A 3D Diffusion Model for Localized Shape Editing

    Rolandos Alexandros Potamias, Michail Tarasiou, Stylianos Ploumpis, and Stefanos Zafeiriou. ShapeFusion: A 3D Diffusion Model for Localized Shape Editing. In Aleˇ s Leonardis, Elisa Ricci, Stefan Roth, Olga Russakovsky, Torsten Sattler, and G¨ ul Varol, editors, Computer Visio...

  21. [29]

    OctFusion: Octree-based Diffusion Models for 3D Shape Generation

    Bojun Xiong, Si-Tong Wei, Xin-Yang Zheng, Yan-Pei Cao, Zhouhui Lian, and Peng-Shuai Wang. OctFusion: Octree-based Diffusion Models for 3D Shape Generation. Computer Graphics Forum, 44(5):e70198, August 2025. 37

  22. [30]

    Recent Advances on Machine Learning for Computational Fluid Dynamics: A Survey, August

    Haixin Wang, Yadi Cao, Zijie Huang, Yuxuan Liu, Peiyan Hu, Xiao Luo, Zezheng Song, Wanjia Zhao, Jilin Liu, Jinan Sun, Shikun Zhang, Long Wei, Yue Wang, Tailin Wu, Zhi-Ming Ma, and Yizhou Sun. Recent Advances on Machine Learning for Computational Fluid Dynamics: A Survey, August

  23. [31]

    Diffusion Models for 3D Generation: A Survey

    Chen Wang, Hao-Yang Peng, Ying-Tian Liu, Jiatao Gu, and Shi-Min Hu. Diffusion Models for 3D Generation: A Survey. Computational Visual Media, 11(1):1–28, 2025

  24. [32]

    Chapelier, R

    M. Chapelier, R. Bouclier, and J.-C. Passieux. Free-Form Deformation Digital Image Correlation (FFD-DIC): A non-invasive spline regularization for arbitrary finite element measurements. Computer Methods in Applied Mechanics and Engineering, 384:113992, 2021

  25. [33]

    Efficient musculoskeletal annotation using free-form deformation

    Norio Fukuda, Shoji Konda, Jun Umehara, and Masaya Hirashima. Efficient musculoskeletal annotation using free-form deformation. Scientific Reports, 14(1):16077, 2024

  26. [34]

    A Novel body-fitted free form deformation method for hull form optimization

    Xinwang Liu, Zitong Rong, Lei Yuan, Xiaohang Ji, Luyao Wang, and Xu Sun. A Novel body-fitted free form deformation method for hull form optimization. Ocean Engineering, 342:122786, 2025

  27. [35]

    A Physics-Informed Deep Learning Deformable Medical Image Registration Method Based on Neural ODEs

    Amirhossein Amiri-Hezaveh, Shelly Tan, Qing Deng, David Umulis, Lau- ren Cunniff, Johannes Weickenmeier, and Adrian Buganza Tepole. A Physics-Informed Deep Learning Deformable Medical Image Registration Method Based on Neural ODEs. International Journal of Computer Vision, 133...

  28. [36]

    Elasticity-based morphing technique and application to reduced-order modeling

    Abbas Kabalan, Fabien Casenave, Felipe Bordeu, Virginie Ehrlacher, and Alexandre Ern. Elasticity-based morphing technique and application to reduced-order modeling. Applied Mathematical Modelling, 141:115929, May 2025

  29. [37]

    On Multivariate Bern- stein Polynomials

    Mama Foupouagnigni and Merlin Mouafo Wouodji´ e. On Multivariate Bern- stein Polynomials. Mathematics, 8(9):1397, August 2020

  30. [38]

    A condition for isotopic approx- imation

    Fr´ ed´ eric Chazal and David Cohen-Steiner. A condition for isotopic approx- imation. Graphical Models, 67(5):390–404, September 2005

  31. [39]

    Hans P. Geering. Optimal Control with Engineering Applications. Springer Berlin Heidelberg, Berlin, Heidelberg, 2007. 38

  32. [40]

    Rozza, D

    G. Rozza, D. B. P. Huynh, and A. T. Patera. Reduced Basis Approximation and a Posteriori Error Estimation for Affinely Parametrized Elliptic Coer- cive Partial Differential Equations. Archives of Computational Methods in Engineering, 15(3):229–275, 2008

  33. [41]

    Pedregosa, G

    F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Pas- sos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. Scikit- learn: Machine Learning in Python. Journal of Machine Learn...

  34. [42]

    Byrd, Mary E

    Richard H. Byrd, Mary E. Hribar, and Jorge Nocedal. An Interior Point Algorithm for Large-Scale Nonlinear Programming. SIAM Journal on Optimization, 9(4):877–900, January 1999

  35. [43]

    Morris W. Hirsch. Differential Topology, volume 33. Springer New York, 1976

  36. [44]

    On the p-norm condition number of the mul- tivariate triangular Bernstein basis

    Tom Lyche and Karl Scherer. On the p-norm condition number of the mul- tivariate triangular Bernstein basis. Journal of Computational and Applied Mathematics, 119(1-2):259–273, July 2000. 39 A Proof of Theorem 2 Proof.By [38], there exists a diffeomorphism fromS 1 toS 2 that i...

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.