Pith. sign in

REVIEW 2 major objections 2 minor 32 references

A neural displacement field on NURBS control points parametrizes the admissible design space for multi-patch CAD surfaces while enabling direct constraint-driven morphing.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.3

2026-06-27 21:08 UTC pith:VVCAIO6L

load-bearing objection The paper combines an MLP displacement field on multi-patch NURBS control points with differentiable hydrostatic integrals for direct CAD optimization, but the abstract supplies no quantitative checks on surface validity or error metrics. the 2 major comments →

arxiv 2606.07198 v1 pith:VVCAIO6L submitted 2026-06-05 math.NA cs.NA

Constraint-driven Optimization and Parametrization of Industrial NURBS Geometries via Neural Deformation Field

Federico Tamburlin , Giovanni Canali , Giuseppe Alessio D'Inverno , Nicola Demo , Andrea Mola , Gianluigi Rozza This is my paper
classification math.NA cs.NA
keywords NURBSshape optimizationneural displacement fielddifferentiable parametrizationhydrostatic constraintsmulti-patch surfacesCAD geometriesGauss-Legendre quadrature
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper presents a differentiable framework that deforms industrial multi-patch NURBS geometries using a multi-layer perceptron to displace control points. This yields a compact representation of design variations that preserves patch connectivity and allows physical constraints such as hydrostatic integrals to drive the shape changes directly. Global quantities including displaced volume, wetted surface area, and buoyancy centroid are expressed as differentiable operators evaluated on the parametric domain via Gauss-Legendre quadrature with analytical B-spline derivatives. The approach operates on the original CAD representation without intermediate meshing, and experiments on a modified KVLCC2 hull show that competing hydrostatic constraints can be satisfied while producing smooth, CAD-compatible results with stable convergence from varied initializations.

Core claim

A neural displacement field, implemented as a multi-layer perceptron acting on the NURBS control points, provides a compact parametrization of the admissible design space while preserving patch connectivity and enables direct morphing driven by physical constraints. Global geometric quantities relevant to hydrostatic design are formulated as differentiable integral operators evaluated on the parametric domain through Gauss-Legendre quadrature combined with analytical B-spline derivatives, allowing gradient propagation to the deformation parameters. The framework operates directly on CAD representations without intermediate mesh generation.

What carries the argument

Neural displacement field (multi-layer perceptron acting on NURBS control points) that supplies a compact parametrization while preserving multi-patch connectivity and supporting differentiable hydrostatic integrals via quadrature.

Load-bearing premise

Displacing NURBS control points via an MLP will always produce valid, non-self-intersecting multi-patch surfaces suitable for industrial CAD use, with the chosen quadrature accurately capturing the global hydrostatic integrals.

What would settle it

A deformed multi-patch NURBS surface generated by the MLP that exhibits self-intersections or produces hydrostatic integral values differing by more than quadrature tolerance from a high-resolution reference evaluation on the same geometry.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Gradient-based optimization can satisfy competing hydrostatic constraints directly on the CAD model.
  • Smooth CAD-compatible geometries result from the optimization with stable convergence across random initializations.
  • Differentiable computation of displaced volume, wetted surface area, and buoyancy centroid is available without mesh generation or predefined deformation maps.
  • Analytical B-spline derivatives limit the overhead of automatic differentiation during gradient propagation.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same MLP parametrization could be applied to other constraint sets such as structural or aerodynamic quantities on NURBS models.
  • Compactness of the learned displacement field may support surrogate modeling or real-time exploration of design spaces in CAD environments.
  • The method could be tested on additional industrial NURBS objects such as aircraft fuselages or turbine blades to check generality beyond ship hulls.
  • Integration with existing CAD kernels might allow the framework to serve as a plug-in optimizer for parametric studies.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 2 minor

Summary. The manuscript introduces a differentiable framework for parametrizing and optimizing multi-patch NURBS CAD geometries via a neural displacement field (MLP acting on control-point coordinates). This enables direct, constraint-driven morphing of industrial models (e.g., a modified KVLCC2 hull) while preserving patch connectivity; global hydrostatic quantities (displaced volume, wetted area, buoyancy centroid) are expressed as differentiable integrals evaluated by Gauss-Legendre quadrature on the parametric domain using analytical B-spline derivatives, allowing gradient-based optimization without intermediate meshing.

Significance. If the generated surfaces remain valid, the approach provides a compact, connectivity-preserving parametrization of the admissible design space that integrates directly with CAD representations. The combination of an MLP deformation field with exact B-spline metric derivatives and quadrature-based differentiation is a technical strength that could facilitate physics-informed shape optimization in naval architecture and related fields.

major comments (2)
  1. [Numerical experiments] Numerical experiments section: the claim that the method produces “smooth CAD-compatible geometries” is load-bearing for the central contribution, yet no quantitative validation is reported (minimum Jacobian determinant over knot spans, minimum inter-patch distance, or knot-span validity tests) to confirm absence of self-intersections, folding, or negative Jacobians under the learned displacements.
  2. [Method formulation] Method formulation (neural displacement field): while absolute-position input to the MLP automatically respects connectivity, the formulation contains no regularization term or projection step that enforces positive surface Jacobians or non-intersection; the abstract’s assertion of validity therefore rests on an unverified assumption rather than an enforced property.
minor comments (2)
  1. Clarify the precise input encoding to the MLP (absolute coordinates versus normalized parameters) and whether the same MLP weights are shared across all patches or per-patch.
  2. The quadrature order and number of quadrature points used for the hydrostatic integrals should be stated explicitly, together with a brief convergence check against a higher-order rule.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which help clarify the scope and limitations of our contribution. We address each major comment below and indicate the corresponding revisions.

read point-by-point responses

  1. Referee: [Numerical experiments] Numerical experiments section: the claim that the method produces “smooth CAD-compatible geometries” is load-bearing for the central contribution, yet no quantitative validation is reported (minimum Jacobian determinant over knot spans, minimum inter-patch distance, or knot-span validity tests) to confirm absence of self-intersections, folding, or negative Jacobians under the learned displacements.

    Authors: We agree that quantitative validation is necessary to substantiate the claim of smooth CAD-compatible geometries. In the revised manuscript we will augment the Numerical experiments section with explicit metrics computed on the final optimized hull: the minimum Jacobian determinant evaluated over all knot spans (using the analytical B-spline derivatives already present in the framework), the minimum inter-patch distance, and a simple knot-span validity check. These quantities will be reported for all random initializations shown in the paper. revision: yes

  2. Referee: [Method formulation] Method formulation (neural displacement field): while absolute-position input to the MLP automatically respects connectivity, the formulation contains no regularization term or projection step that enforces positive surface Jacobians or non-intersection; the abstract’s assertion of validity therefore rests on an unverified assumption rather than an enforced property.

    Authors: We acknowledge that the neural displacement field is formulated without an explicit regularization term or projection that guarantees positive Jacobians or non-intersection; validity is therefore observed empirically rather than enforced by construction. The absolute-position input ensures only connectivity preservation. In the revised manuscript we will clarify this distinction in the Method section and add a short discussion of possible future extensions (e.g., a Jacobian-penalty term in the loss), while retaining the current formulation as a compact, connectivity-preserving parametrization whose practical validity is demonstrated by the reported experiments. revision: partial

Circularity Check

0 steps flagged

No circularity: MLP parametrization and constraint optimization are independent

full rationale

The derivation introduces an MLP-based neural displacement field as a modeling choice to parametrize NURBS control-point displacements while preserving connectivity by construction of the shared MLP. Hydrostatic quantities (volume, wetted area, buoyancy centroid) are then defined as independent differentiable integrals evaluated via Gauss-Legendre quadrature on the parametric domain. These external physical targets drive optimization of the MLP weights; the resulting geometry is not equivalent to any fitted input by definition, nor does any step rename a known result or rely on self-citation chains. The framework remains self-contained against the stated constraints without reduction to tautology.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The framework rests on standard properties of NURBS surfaces and the ability of an MLP to generate admissible deformations; no new physical entities are introduced.

free parameters (1)
  • MLP weights and biases
    Parameters of the multi-layer perceptron that define the displacement field are optimized during the constraint-driven process.
axioms (2)
  • domain assumption Displacing NURBS control points via a learned field preserves patch connectivity and produces valid CAD geometries.
    Invoked when stating that the neural field provides a compact parametrization while preserving connectivity.
  • domain assumption Gauss-Legendre quadrature on the parametric domain combined with analytical B-spline derivatives yields sufficiently accurate gradients for the hydrostatic integrals.
    Stated as the mechanism allowing gradient propagation without excessive computational overhead.

pith-pipeline@v0.9.1-grok · 5745 in / 1529 out tokens · 18371 ms · 2026-06-27T21:08:13.291078+00:00 · methodology

0 comments
read the original abstract

This work presents a differentiable framework for the parametrization and shape optimization of industrial CAD geometries represented by multi-patch NURBS surfaces. The method enables the deformation of complex CAD models through a physics-informed geometric parametrization, allowing direct morphing driven by physical constraints without the need to prescribe a predefined deformation strategy. A neural displacement field, implemented as a multi-layer perceptron acting on the NURBS control points, provides a compact parametrization of the admissible design space while preserving patch connectivity. Global geometric quantities relevant to hydrostatic design, including displaced volume, wetted surface area and buoyancy centroid, are formulated as differentiable integral operators evaluated on the parametric domain. These quantities are computed through Gauss-Legendre quadrature combined with analytical B-spline derivatives for surface metric evaluation, allowing gradient propagation to the deformation parameters while limiting the computational overhead of automatic differentiation. The proposed framework operates directly on CAD representations without intermediate mesh generation. Numerical experiments on a modified KVLCC2 hull demonstrate the capability of the method to satisfy competing hydrostatic constraints while producing smooth CAD-compatible geometries and showing stable convergence across multiple random initializations.

Figures

Figures reproduced from arXiv: 2606.07198 by Andrea Mola, Federico Tamburlin, Gianluigi Rozza, Giovanni Canali, Giuseppe Alessio D'Inverno, Nicola Demo.

Figure 1
Figure 1. Figure 1: Schematic of the end-to-end differentiable optimization pipeline. The system maps original geometrical control points through a Neural Displacement Field (MLP) to generate the deformed surface. Geometric properties (Volume, Area, Centroid) are computed via analytical B-Spline derivatives and numerical quadrature, driving the loss function composed of integral constraints, geometric barrier and regularizati… view at source ↗
Figure 2
Figure 2. Figure 2: Neural Deformation architecture: the displacement is computed by the MLP and added to the original control points. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Convergence and stability analysis of geometric integrals via Gauss-Legendre quadrature. [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Optimization convergence analysis (Seed= 911 and λreg = 1e −4 ). The plots illustrate the evolution of the global loss and specific geometric constraint losses against the number of function evaluations. Raw oscillation data is shown with high transparency, overlaid by a solid line representing a moving average to enhance trend visibility. The vertical dashed line indicates the switch from the stochastic e… view at source ↗
Figure 5
Figure 5. Figure 5: Visualization of the optimized geometric deformation. Overlay of the initial hull (blue) and the optimized hull (red). Note the significant [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparative cross-sectional analysis of the modified KVLCC2 hull geometry optimized via Free-Form Deformation (FFD) and the [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Impact of the parametrization strategy on geometric quality. The direct optimization of control points lacks regularization, resulting in [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Comparative cross-sectional analysis at Station 10 and Station 18 over 20 total sections. The negative [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Robustness analysis of the neural parametrization with respect to stochastic initialization. Comparison of cross-sections at various [PITH_FULL_IMAGE:figures/full_fig_p016_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Sensitivity analysis of the geometric optimization performance with respect to the regularization weight [PITH_FULL_IMAGE:figures/full_fig_p017_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Comparison of optimisation dynamics for two representative regularization regimes (Seed 911). The vertical dashed line marks the switch [PITH_FULL_IMAGE:figures/full_fig_p018_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Sensitivity of geometric deformation to regularization stiffness. Cross-sections at four representative stations. (c) At midship, the Low Stiffness regime (orange, λreg = 10−5 ) enables the necessary beam expansion to compensate for the draft reduction (Zmin constraint), ensuring volume conservation within numerical tolerance. In contrast, the High Stiffness regime (red, λreg = 10−2 ) locks the geometry t… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

32 extracted references · 5 canonical work pages · 1 internal anchor

  1. [1]

    Srinivasan, An integration framework for product lifecycle management, Computer-aided design 43 (5) (2011) 464–478

    V . Srinivasan, An integration framework for product lifecycle management, Computer-aided design 43 (5) (2011) 464–478

    2011

  2. [2]

    Sudarsan, S

    R. Sudarsan, S. J. Fenves, R. D. Sriram, F. Wang, A product information modeling framework for product lifecycle management, Computer-aided design 37 (13) (2005) 1399–1411

    2005

  3. [3]

    Z. Bi, X. Wang, Computer aided design and manufacturing, John Wiley & Sons, 2020

    2020

  4. [4]

    K. J. Versprille, Computer-aided design applications of the rational b-spline approximation form., Syracuse University, 1975

    1975

  5. [5]

    Ivagnes, N

    A. Ivagnes, N. Demo, G. Rozza, A shape optimization pipeline for marine propellers by means of reduced order modeling techniques, International Journal for Numerical Methods in Engineering (2024). arXiv:2305.07515, doi:10.1002/nme.7426. 19

    work page doi:10.1002/nme.7426 2024

  6. [6]

    N. Demo, M. Tezzele, A. Mola, G. Rozza, Hull shape design optimization with parameter space and model reductions, and self-learning mesh morphing, Journal of Marine Science and Engineering 9 (2) (2021). arXiv: 2101.03781,doi:10.3390/jmse9020185. URLhttps://www.mdpi.com/2077-1312/9/2/185

    work page doi:10.3390/jmse9020185 2021

  7. [7]

    Villa, F

    D. Villa, F. Furcas, J. O. Pralits, G. Vernengo, S. Gaggero, An effective mesh deformation approach for hull shape design by optimization, Journal of Marine Science and Engineering 9 (10) (2021) 1107

    2021

  8. [8]

    T. W. Sederberg, S. R. Parry, Free-form deformation of solid geometric models, in: Proceedings of the 13th annual conference on Computer graphics and interactive techniques, 1986, pp. 151–160

    1986

  9. [9]

    D. Wei, L. Wang, J. Yang, X. Sun, X. Ji, X. Liu, Hull form optimization with a new three-dimensional deformation strategy, Scientific Reports 15 (1) (2025) 41804

    2025

  10. [10]

    S. Li, F. Zhu, X. Hou, Q. Ni, Application of mesh deformation and adaptive method in hullform design optimiza- tion, Journal of Marine Science and Technology 27 (1) (2022) 566–575

    2022

  11. [11]

    S. Wang, Y . Kim, Cfd-based hull optimization in calm water using adaptive grid deformation method, Journal of Hydrodynamics 36 (2) (2024) 241–262

    2024

  12. [12]

    A. Mola, N. Demo, M. Tezzele, G. Rozza, Chapter 17: Geometrical Parametrization and Morphing Techniques with Applications, 2022, pp. 345–364.doi:10.1137/1.9781611977257.ch17. URLhttps://epubs.siam.org/doi/abs/10.1137/1.9781611977257.ch17

    work page doi:10.1137/1.9781611977257.ch17 2022

  13. [13]

    J. Lee, S. Hube, S. Elgeti, Neural networks vs. splines: advances in numerical extruder design, Engineering with Computers 40 (2) (2024) 989–1004

    2024

  14. [14]

    S. Khan, K. Goucher-Lambert, K. Kostas, P. Kaklis, Shiphullgan: A generic parametric modeller for ship hull design using deep convolutional generative model, Computer Methods in Applied Mechanics and Engineering 411 (2023) 116051

    2023

  15. [15]

    N. J. Bagazinski, F. Ahmed, Shipgen: A diffusion model for parametric ship hull generation with multiple objectives and constraints, Journal of Marine Science and Engineering 11 (12) (2023) 2215

    2023

  16. [16]

    Padula, F

    G. Padula, F. Romor, G. Stabile, G. Rozza, Generative models for the deformation of industrial shapes with linear geometric constraints: Model order and parameter space reductions, Computer Methods in Applied Mechanics and Engineering 423 (2024) 116823

    2024

  17. [17]

    Raissi, P

    M. Raissi, P. Perdikaris, G. E. Karniadakis, Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations, Journal of Computational physics 378 (2019) 686–707

    2019

  18. [18]

    Y . Sun, U. Sengupta, M. Juniper, Physics-informed deep learning for simultaneous surrogate modeling and PDE-constrained optimization of an airfoil geometry, Computer Methods in Applied Mechanics and Engineering 411 (2023) 116042

    2023

  19. [19]

    Jalui, J

    S. Jalui, J. O’Connor, Y . Xuan, G. Manogharan, Developing novel restrictive design for additive manufacturing (DfAM) constraints for NURBS-based adjoint shape optimization for metal AM, Graphical Models 142 (2025) 101304

    2025

  20. [20]

    Mohebbi, B

    F. Mohebbi, B. Evans, On an exact step length in gradient-based aerodynamic shape optimization, Fluids 5 (2) (2020) 70

    2020

  21. [21]

    Moola, A

    A. Moola, A. Balu, A. Krishnamurthy, A. Pawar, THB-Diff: a GPU-accelerated differentiable programming framework for THB-splines, Engineering with Computers 40 (6) (2024) 3477–3493

    2024

  22. [22]

    Tillmann, D

    S. Tillmann, D. Hilger, N. Hosters, S. Elgeti, Shape-optimization of extrusion-dies via parameterized physics- informed neural networks, PAMM 23 (4) (2023) e202300203. 20

    2023

  23. [23]

    Zhang, C

    Z. Zhang, C. Lin, B. Wang, Physics-informed shape optimization using coordinate projection, Scientific Reports 14 (1) (2024) 6537

    2024

  24. [24]

    Coscia, A

    D. Coscia, A. Ivagnes, N. Demo, G. Rozza, Physics-informed neural networks for advanced modeling, Journal of Open Source Software 8 (87) (2023) 5352

    2023

  25. [25]

    Neural reparameterization improves structural optimization

    S. Hoyer, J. Sohl-Dickstein, S. Greydanus, Neural reparameterization improves structural optimization, arXiv preprint arXiv:1909.04240 (2019)

    work page Pith review arXiv 1909

  26. [26]

    Chandrasekhar, K

    A. Chandrasekhar, K. Suresh, TOuNN: topology optimization using neural networks, Structural and Multidisci- plinary Optimization 63 (3) (2021) 1135–1149

    2021

  27. [27]

    Piegl, W

    L. Piegl, W. Tiller, The NURBS book, Springer Science & Business Media, 2012

    2012

  28. [28]

    Abramowitz, Handbook of mathematical functions with formulas, graphs, and mathematical tables, nbs, Applied Math

    M. Abramowitz, Handbook of mathematical functions with formulas, graphs, and mathematical tables, nbs, Applied Math. Series 55 (1964) 887

    1964

  29. [29]

    CAD source: modified KVLCC2 and modified pyiges, https://github.com/fedetamburlin/ cad-pinn-souces(2026)

    2026

  30. [30]

    Rahaman, A

    N. Rahaman, A. Baratin, D. Arpit, F. Draxler, M. Lin, F. Hamprecht, Y . Bengio, A. Courville, On the spectral bias of neural networks, in: International Conference on Machine Learning, PMLR, 2019, pp. 5301–5310

    2019

  31. [31]

    Decoupled Weight Decay Regularization

    I. Loshchilov, F. Hutter, Decoupled weight decay regularization, arXiv preprint arXiv:1711.05101 (2017)

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  32. [32]

    D. R. S. Saputro, P. Widyaningsih, Limited memory Broyden-Fletcher-Goldfarb-Shanno (L-BFGS) method for the parameter estimation on geographically weighted ordinal logistic regression model (GWOLR), in: AIP conference proceedings, V ol. 1868 (1), AIP Publishing LLC, 2017, p. 040009. 21

    2017