Pith. sign in

REVIEW 4 major objections 5 minor 28 references

A dual diffusion model trained on raw 3D blade coordinates generates centrifugal-compressor geometries that hit prescribed mass flow, speed, pressure ratio and efficiency.

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 ↗

Conditional diffusion models can invert meanline compressor performance into diverse, feasible 3D blade geometries with sub-percent error against that same meanline surrogate.

T0 review reviewed 2026-07-30 challenge →

load-bearing objection Solid first demo of direct-coordinate 3D diffusion for compressor inverse design; the ~99% accuracy numbers are almost entirely meanline-closed and the abstract overclaims transfer. the 4 major comments →

arxiv 2607.27093 v1 pith:HAD5VCII submitted 2026-07-29 physics.flu-dyn

A New Paradigm for 3D Turbomachinery Design: Generative Diffusion Model Based Framework with Direct Geometry Encoding

Yingfan Geng , Jinhong Wang , Lazaros Papachristodoulou , Sibo Cheng , Teng Cao This is my paper
classification physics.flu-dyn
keywords diffusion modelaerodynamic designturbomachinerygenerative AI for designcentrifugal compressorinverse design3D geometry encoding
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 reading

Turbomachinery blades are hard to design because flow physics is nonlinear and many geometries can meet the same performance targets. This paper shows that a conditional denoising diffusion model can solve the inverse problem directly: given target mass-flow rate, rotational speed, pressure ratio and efficiency, it returns full 3D compressor blade coordinates. Unlike earlier generative work that relied on hand-chosen parameters, the main network is trained on normalised cylindrical point grids (3 channels × 16 spans × 512 points) with an auxiliary network that supplies the bounding-box and blade-count descriptors. Against a meanline evaluator the designs reach roughly half-percent RMSE on pressure ratio and efficiency, fewer than one percent of samples are physically invalid, and the cloud of solutions covers the design space as thoroughly as Latin-hypercube sampling of the original parameters. Once trained, the same model can be re-sampled to produce a pool of candidates that all satisfy the primary targets and can then be ranked by any secondary criterion without retraining.

Core claim

A dual Elucidated Diffusion Model architecture, conditioned on operating point and performance targets and trained directly on locally-normalised 3D blade coordinates rather than parametric descriptions, produces centrifugal-compressor geometries whose meanline-evaluated pressure ratio and efficiency match the prescribed targets to within ~0.5 % RMSE, with failure rates below 1 % and solution diversity statistically comparable to physical-parameter sampling.

What carries the argument

The dual diffusion pair: an auxiliary MLP that samples the seven-dimensional geometry-descriptor vector (bounding-box extents plus blade count) and a U-Net main model that denoises the full 3 imes16 imes512 normalised cylindrical coordinate tensor conditioned on that descriptor plus the four aerodynamic targets.

Load-bearing premise

The paper treats meanline-model pressure ratio and efficiency, built from empirical loss correlations on simplified full-blade impellers, as a faithful enough stand-in for true aerodynamic performance that accuracy measured against the same meanline transfers to physically valid designs.

What would settle it

Generate a statistically meaningful set of blades from the trained 3D model, evaluate them with high-fidelity CFD (or experiment) at the prescribed operating points, and check whether pressure-ratio and efficiency errors remain near the reported half-percent level and whether the fraction of aerodynamically invalid geometries stays below one percent.

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

If this is right

  • Inverse design of centrifugal compressors can be reduced to a single conditioned sampling pass that returns multiple feasible 3D geometries.
  • Secondary objectives (off-design maps, thickness constraints, manufacturing limits) can be imposed by post-filtering the sampled pool without retraining.
  • The same coordinate-based pipeline can be applied to blade families that lack convenient parametric descriptions, such as bio-inspired or free-form shapes.
  • Once higher-fidelity performance labels replace the meanline, the identical architecture becomes a drop-in generator for CFD-validated designs.

Where Pith is reading between the lines

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

  • If the dual-model split is unnecessary once enough high-fidelity data exist, a single end-to-end diffusion network on raw coordinates could further simplify the workflow.
  • The same local-normalisation-plus-descriptor pattern should transfer to other rotational machines (axial compressors, turbines, pumps) whose geometry lives naturally in cylindrical coordinates.
  • Coupling the sampler to a differentiable CFD surrogate would let secondary aerodynamic targets be folded into the reverse-diffusion guidance itself rather than applied only as a filter.
Share X Bluesky LinkedIn Reddit HN

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

4 major / 5 minor

Summary. The paper proposes a dual conditional diffusion (EDM) framework for inverse aerodynamic design of centrifugal compressor impellers. A main U-Net diffusion model generates normalised 3D blade coordinates (3×16×512 in cylindrical polar form), conditioned on mass flow, rotational speed, pressure ratio, efficiency, and a geometry-descriptor vector g; an auxiliary MLP diffusion model can supply g automatically. A 1D-parameter baseline is trained for comparison. Training and performance labels come from Latin-hypercube sampling of eight design variables, an in-house blade former, and a calibrated meanline model with empirical loss correlations. On a held-out test set the models report PR/η RMSE of order 0.3–0.5% versus the same meanline evaluator, failure rates below ~2%, and solution-set diversity (Chamfer, SSIM, JSD) comparable to independent LHS of the physical parameter space. One generated geometry is checked with 3D RANS CFD.

Significance. Direct training on structured 3D blade coordinates rather than hand-chosen parameters is a genuine methodological step for turbomachinery generative design, and the dual main/auxiliary architecture with optional manual g is a clean way to keep geometric constraints. The EDM formulation, local cylindrical normalisation, diversity metrics against physical LHS, and multi-target filtering procedure are clearly specified and reproducible in principle. If the accuracy and feasibility claims transfer beyond the meanline surrogate, the work would offer a practical inverse-design paradigm. As written, the strongest supported contribution is the geometry-generation and design-space exploration framework; the aerodynamic ‘~99% accuracy’ claim remains tied to the labelling surrogate and needs firmer independent evidence.

major comments (4)
  1. [Abstract; §2.4.1; §3.1; Table 4] Abstract and §3.1 / Table 4 report solution accuracy ‘up to 99%’ and PR/η RMSE ≈ 0.33–0.53% with failure <1% (3D). These figures are obtained by extracting 1D parameters from generated 3D surfaces and re-evaluating with the identical meanline model that produced the training labels (§2.2.2, §2.4.1, Eqs. 16–17). That loop shows the diffusion models have learned the conditional geometry distribution of the surrogate; it does not by itself establish physical inverse-design accuracy. The abstract and conclusions should state this evaluation basis explicitly, and the central accuracy claim should be qualified accordingly.
  2. [§3.4; Figure 13] Independent aerodynamic evidence is limited to a single CFD map on one design from the multi-target pool (§3.4). At the design point the CFD PR already differs by ~2% from the specified target (η by ~0.2%), larger than the meanline RMSE in Table 4. One case cannot support transfer of the reported accuracy/diversity results. At minimum, CFD (or another independent evaluator) on a small stratified set of generated designs—including 1D vs 3D and near-failure cases—should be added, or the claim scope narrowed to ‘meanline-consistent inverse design’ with CFD as a feasibility check only.
  3. [§2.5.2; §3.2; Table 5] Diversity (§2.5.2, §3.2, Table 5) compares diffusion solution sets A to LHS sets B1/B2 that are themselves filtered by the same meanline tolerance on PR and η. CD/SSIM/JSD therefore measure agreement of geometry distributions under a shared surrogate filter, not necessarily coverage of the true physical manifold. This is still useful, but the text should not equate it with ‘exploration of the entire design space’ without noting the shared evaluator. If 3D↔1D extraction or meanline bias systematically mis-ranks shapes, both accuracy and diversity metrics move together.
  4. [§2.4.1; §3.1] Performance evaluation of 3D outputs requires ‘measurements… to extract the 1D geometry parameters’ plus optional polynomial smoothing (~0.1% of axial length) before the meanline call (§2.4.1, §3.1). The manuscript does not quantify how sensitive PR/η are to that extraction and smoothing, nor whether extraction error differs between 1D-baseline blades (formed by the same tool) and 3D-diffusion blades (pointwise noise then fitted). A short sensitivity or consistency check is needed so that the 1D vs 3D comparison in Table 4 is not confounded by post-processing.
minor comments (5)
  1. [§2.2.3; Table 1–2; Fig. 4] Table 1 lists inlet tip radius as R_tip,1 but the text under Fig. 4 sometimes equates r_min with R_tip,1 and sometimes with hub radius; align notation with Table 2 (R_hub,1 / R_tip,1).
  2. [§2.3.1; Eq. (8); Algo. 1] Eq. (8) mixes ± on the SDE term in a way that is hard to parse; a short clarification that the ODE path is used with optional Langevin noise (as in Algo. 1) would help readers not steeped in EDM.
  3. [Fig. 3; Fig. 5; Fig. 10] Fig. 3 and Fig. 5 KDEs lack axis units/bandwidth; Fig. 10 scatter panels would benefit from explicit RE colouring or tolerance bands matching ε = 1%.
  4. [§1] Related work on diffusion for turbomachinery (§1) is adequate but could briefly contrast conditioning choices (e.g. pressure-coefficient distributions in [5,6]) with the operating-point + PR/η conditioning used here.
  5. [Fig. 1; throughout] Typos/wording: ‘Normalsied’, ‘auxliary’, ‘consitions’, ‘unfeasible’ vs ‘infeasible’, and arXiv date ‘29 Jul 2026’ look like residual draft artefacts.

Circularity Check

1 steps flagged

No derivation-by-construction circularity; meanline-closed accuracy is standard surrogate evaluation, not a tautology.

specific steps
  1. self citation load bearing [§2.3.1; also Introduction contributions and [8]]
    "A conditional diffusion model serves as the generative backbone of this work, following the work in [8]. ... Following the work on 2D aerofoil designs in [8], this paper introduces a diffusion model-based 3D turbomachinery inverse design method"

    The EDM sampling procedure, single-/multi-target generation loops, and evaluation style are imported from the authors’ own prior aerofoil paper [8]. This is methodological continuity, not a uniqueness theorem that forces the compressor accuracy numbers; the 3D dual-model results are still trained and measured on new data. Minor and not load-bearing for the numerical claims—hence only a +1, not a high score.

full rationale

The paper’s load-bearing claim is that conditional EDMs (main U-Net on normalised 3×16×512 blade coordinates plus auxiliary MLP on geometry descriptors; 1D MLP baseline) sample geometries whose meanline-evaluated PR and η match prescribed targets within ~0.3–0.5% RMSE, with low failure rate and diversity comparable to LHS of the physical parameter space. Training labels and the accuracy metric both use the same meanline model (§2.2.2, §2.4.1, Eqs. 16–17, Table 4). That is a closed evaluation loop relative to high-fidelity physics, but it is not circularity in the sense of this pass: accuracy is measured on held-out target conditions, and a failed denoiser would yield large RMSE—the reported numbers are not identities or forced fits. Diversity (Table 5, Fig. 11) compares diffusion solution sets A to independently LHS-sampled sets B1/B2 in the same reduced parameter space; CD/SSIM/JSD agreement is empirical, not definitional. Self-citations ([8] prior 2D aerofoil EDM workflow; [15] meanline tool with author overlap) supply method and the surrogate; neither imports a uniqueness theorem nor makes the accuracy/diversity results true by citation. Single CFD map (§3.4) is an external check (and already shows ~2% PR offset), confirming the main tables are meanline-consistency metrics. No equation equates a claimed prediction to its fitted input by construction. Score 1 only for the mild, non-load-bearing self-citation of the authors’ own prior diffusion design pipeline.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 1 invented entities

The central inverse-design claim rests on standard diffusion math, domain engineering assumptions (meanline adequacy, geometry manifold, LHS coverage), and several paper-specific modelling choices (dual-model split, local cylindrical normalisation, tolerance/trial protocol). No new physical entities; free parameters are mostly training/sampling hyperparameters and the empirical meanline loss package treated as ground truth.

free parameters (5)
  • Performance acceptance tolerance ε (PR and η) = 1%
    Fixed at 1% for declaring a generated design valid; directly controls reported trial counts and success framing.
  • Max sampling trials N_trial = 100
    Capped at 100; affects average trial metric and failure handling when ε is unmet.
  • EDM / network training hyperparameters = EDM defaults; main model 300 epochs, batch 32
    Default EDM noise schedule and solver settings; U-Net/MLP depth, width, epochs (100/300), batch sizes; chosen rather than derived.
  • Meanline empirical loss-model coefficients = As in cited meanline calibration [15]
    Incidence, leakage, clearance, shock, choke, friction losses calibrated elsewhere; all PR/η labels and most evaluation depend on them.
  • Geometry discretisation (16 spans × 512 points) = 3×16×512
    Fixed structured grid that defines the 3D learning space and Chamfer/SSIM metrics.
axioms (6)
  • standard math Elucidated Diffusion Model probability-flow ODE/SDE and score-matching training objective are valid generative models for this geometry distribution.
    Adopted from Karras et al. EDM; §2.3.1 Eqs. 8–10 and Algo. 1.
  • domain assumption Meanline thermodynamic + empirical loss models supply sufficiently accurate PR and η for training labels and accuracy assessment of subsonic centrifugal impellers in the sampled range.
    Stated as proof-of-concept choice in §2.2.2; underpins all quantitative accuracy claims.
  • domain assumption Feasible compressor geometries for fixed (ṁ, ω, PR, η) lie on a learnable low-dimensional manifold in the high-dimensional coordinate space.
    Invoked in §2.4.1 and §2.5.2 to justify generative coverage and diversity metrics.
  • domain assumption Latin-hypercube sampling of eight initial variables plus the in-house blade former adequately spans the relevant design space (~2000 valid shapes).
    §2.2.1 Table 1; geometrically infeasible samples discarded.
  • ad hoc to paper Local per-blade min-max normalisation in cylindrical (x,r,θ) plus a separate global-normalised descriptor vector is a faithful, invertible geometry encoding.
    §2.2.3 Eqs. 5–7 and dual-model conditioning Eqs. 11–14; enables direct coordinate learning but is a paper-specific representation choice.
  • ad hoc to paper Extracting 1D meanline inputs from generated 3D surfaces (and light polynomial smoothing) does not materially corrupt performance evaluation.
    §2.4.1 and §3.1; required to close the 3D-generation vs meanline loop.
invented entities (1)
  • Dual main+auxiliary diffusion architecture for 3D blades with geometry descriptor vector g no independent evidence
    purpose: Separate global scale/blade-count generation from dense normalised surface generation; support automatic vs manual constraint modes.
    Introduced in §2.3.2 Eqs. 11–14 and Fig. 1; engineering architecture, not a new physical object.

reviewed 2026-07-30 · how reviews work

0 comments
Cite this review

Pith. "Pith review of A New Paradigm for 3D Turbomachinery Design: Generative Diffusion Model Based Framework with Direct Geometry Encoding." pith.science (2026). https://pith.science/paper/HAD5VCII

@misc{pith2026260727093,
  author       = {Pith},
  title        = {Pith review of: A New Paradigm for 3D Turbomachinery Design: Generative Diffusion Model Based Framework with Direct Geometry Encoding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HAD5VCII}},
  note         = {Machine review of arXiv:2607.27093}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

The aerodynamic design of turbomachinery is critical to the performance of the overall energy system, yet it is challenging due to the complex non-linear flow physics and the presence of multiple-target design compromises. Denoising diffusion model, as one of the leading approaches in generative machine learning, has shown its advantages of high design solution accuracy and diversity in many engineering applications. In this study, we bring it to the 3D inverse design problem of turbomachinery, using centrifugal compressors as a classic representative, to demonstrate the new methodology for complex geometry designs. A diffusion model-centred design framework has been developed in this study. By specifying the desired design condition (mass flow rate and rotational speed) and targeted performance (pressure ratio and efficiency), the trained diffusion model returns directly the 3D compressor geometry that satisfies the condition inputs. Compared to traditional deterministic forward design approaches, the proposed method not only generates accurate geometry solutions to inverse design problems, but also enables effective exploration of the entire design space, providing a diverse set of candidate solutions. In addition, this paper presents the first study to directly train on 3D blade geometry coordinates rather than parametrised representations, demonstrating the feasibility of coordinate-based learning while enabling a highly flexible framework applicable to a wide range of designs. The trained diffusion model achieves excellent design capability, with solution accuracy up to 99% and unfeasible designs less than 1%. Furthermore, the solution diversity of the trained diffusion model is also quantitatively verified by means of comparing the distribution of the solution sets generated from the diffusion model and from direct sampling of physical parameters.

Figures

Figures reproduced from arXiv: 2607.27093 by Jinhong Wang, Lazaros Papachristodoulou, Sibo Cheng, Teng Cao, Yingfan Geng.

Figure 1
Figure 1. Figure 1: Proposed Centrifugal Compressor Inverse Design Framework with Diffusion Model 2. Methodology 2.1. Proposed Design Framework The overall design framework of this study is illustrated by [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Main Impeller Geometry [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Compressor Performance Database Distribution (a) Cartesian (𝑥, 𝑦, 𝑧) (b) Cylindrical polar (𝑥, 𝑟, 𝜃) (c) Normalised cylindrical polar (𝑥, 𝑟, 𝜃) [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: 3D compressor blade geometry presentation in different coordinate systems. To make them visually distinguishable, the blade in the normalised coordinate is shown with discretised points in blue. 𝑥, 𝑦, 𝑧 coordinates into cylindrical polar coordinates (𝑥, 𝑟, 𝜃) as it better aligns with the intrinsic geometric and rotational periodic features of the configuration. Figure 4a presents a compressor blade in 𝑥, 𝑦… view at source ↗
Figure 5
Figure 5. Figure 5: Kernel density estimate of the normalised compressor geometry distribution (normalised meridional view 𝑟 vs. 𝑥) length 𝐿𝑧 in [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: UNet Architecture for Denoising Neural Network In summary, the overall blade design generation in [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Proposed Design Generation Procedures of Diffusion Model inputs and from the ground truth calculated using the meanline model (Section 2.4.1), respectively. RMSEPR = √√√√√ 1 𝑁test 𝑁 ∑test 𝑖=1 ( PRactual𝑖 − PRtarget𝑖 PRtarget𝑖 )2 RMSE𝜂 = √√√√√ 1 𝑁test 𝑁 ∑test 𝑖=1 ( 𝜂actual𝑖 − 𝜂target𝑖 𝜂target𝑖 )2 (17) It is to note that the metrics presented by Eq. (17) provide an overall indication of model performance ove… view at source ↗
Figure 8
Figure 8. Figure 8: Denoising Process of the 3D Main Diffusion Model [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Examples of Generated Designs [PITH_FULL_IMAGE:figures/full_fig_p016_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Design Accuracy Plot for Single-Target Generation model deployment is more complicated due to the dual diffusion model architecture and the 3D-to-1D post-processing process, which adds additional uncertainties during accuracy evaluation. Nevertheless, the 3D diffusion model still shows promising results with high design solution accuracies, proving the concept of directly applying the diffusion model to 3… view at source ↗
Figure 11
Figure 11. Figure 11: The KDE Plots of the Four Solution Sets on the Meridional Plane [PITH_FULL_IMAGE:figures/full_fig_p018_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Multi-Targets Generation Results they exhibit significantly different off-design performances, and the best-performing profile can be selected according to the particular application requirement, such as maximising operating range. In fact, once trained, this selection can be flexibly extended to any targets, both geometrically and aerodynamically, without the need to retrain the diffusion model. 3.4. 3D … view at source ↗
Figure 13
Figure 13. Figure 13: Compressor maps produced from the CFD simulations 2 % for pressure ratio and by 0.2 % for efficiency. This performance deviation is within expectation and is considered acceptable, given the inherent uncertainties of the diffusion and meanline models. Overall, the CFD analysis validates the quality and physical validity of the compressor geometry generated through the framework presented in the current st… 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

28 extracted references · 12 canonical work pages · 1 internal anchor

  1. [1]

    P. M. Came, C. J. Robinson, Centrifugal compressor design, Proceedings of the Institution of Mechanical Engineers, Part C: Journal of Mechanical Engineering Science 213 (1998) 139–155. doi:10.1243/0954406991522239

    work page doi:10.1243/0954406991522239 1998

  2. [2]

    Z. Zou, P. Xu, Y. Chen, L. Yao, C. Fu, Application of artificial intelligence in turbomachinery aerodynamics: Progresses and challenges, Artificial Intelligence Review 57 (2024) 222. doi:10.1007/s10462-024-10867-3

    work page doi:10.1007/s10462-024-10867-3 2024

  3. [3]

    Cutrina Vilalta, H

    P. Cutrina Vilalta, H. Wan, S. S. Patnaik, Modeling of Centrifugal Compressor Performance Using Machine Learning Techniques, in: AIAA Scitech 2020 Forum, American Institute of Aeronautics and Astronautics, Orlando, FL, 2020, pp. 1–8. doi:10.2514/6.2020-0134

    work page doi:10.2514/6.2020-0134 2020

  4. [4]

    Li,C.Zhang, Y.Li, Machinelearning baseddesignoptimization ofcentrifugal impellers, Journal oftheGlobal Power and Propulsion Society 6 (2022) 124–134

    A.Zhang, Y.Liu, J.Yang,Z. Li,C.Zhang, Y.Li, Machinelearning baseddesignoptimization ofcentrifugal impellers, Journal oftheGlobal Power and Propulsion Society 6 (2022) 124–134. doi:10.33737/jgpps/150663

    work page doi:10.33737/jgpps/150663 2022

  5. [5]

    doi:10.1115/1.4068386

    X.Jia,X.Huang,S.Jiang,D.Wang,C.M.Firrone, InverseDesignofCompressor/FanBladeProfilesBasedonConditionalInvertibleNeural Networks, Journal of Turbomachinery 147 (2025). doi:10.1115/1.4068386

    work page doi:10.1115/1.4068386 2025

  6. [6]

    doi:10.1115/1.4052301

    S.Ghosh,G.AnanthaPadmanabha,C.Peng,V.Andreoli,S.Atkinson,P.Pandita,T.Vandeputte,N.Zabaras,L.Wang, InverseAerodynamic Design of Gas Turbine Blades Using Probabilistic Machine Learning, Journal of Mechanical Design 144 (2021). doi:10.1115/1.4052301

    work page doi:10.1115/1.4052301 2021

  7. [7]

    J. Ho, A. Jain, P. Abbeel, Denoising Diffusion Probabilistic Models, 2020. doi:10.48550/arXiv.2006.11239.arXiv:2006.11239

    Pith/arXiv arXiv doi:10.48550/arxiv.2006.11239 2020

  8. [8]

    Y.Geng,J.Wang,T.Cao,DiffusionModelDrivenAirfoilDesign:FromGeometryEncodingtoPracticalApplications,2026.doi:10.48550/ arXiv.2601.16228.arXiv:2601.16228

    work page doi:10.48550/arxiv.2601.16228 2026

  9. [9]

    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 (2023) 2215. doi:10.3390/jmse11122215

    work page doi:10.3390/jmse11122215 2023

  10. [10]

    doi:10.1016/j.ast.2026.112049

    K.Chen,C.Qiu,Z.Liu,J.Tan, Anovelperformance-geometrycollaboratedgenerativedesignframeforcomplexsurfaceproduct, Aerospace Science and Technology 176 (2026) 112049. doi:10.1016/j.ast.2026.112049

    arXiv 2026

  11. [11]

    doi:10.1063/5.0266874

    Y.Wu,X.Xu,M.Zhang,D.Ba,J.Du, Aninverseaerodynamicdesignframeworkforcompressorbladesbasedongenerativemodel, Physics of Fluids 37 (2025) 076108. doi:10.1063/5.0266874

    work page doi:10.1063/5.0266874 2025

  12. [12]

    doi:10.1115/1.4069801

    Y.Wu,M.Zhang,X.Xu,Y.Yin,D.Ba,J.Du,GenerativeInverseAerodynamicDesignofaSingle-StageTurbineUsingConditionalDenoising Diffusion Probabilistic Model, Journal of Turbomachinery 148 (2025). doi:10.1115/1.4069801

    work page doi:10.1115/1.4069801 2025

  13. [13]

    Karras, M

    T. Karras, M. Aittala, T. Aila, S. Laine, Elucidating the Design Space of Diffusion-Based Generative Models, 2022. doi:10.48550/arXiv. 2206.00364.arXiv:2206.00364

    Pith/arXiv arXiv doi:10.48550/arxiv.2206.00364 2022

  14. [14]

    M. D. Mckay, R. J. Beckman, W. J. Conover, A Comparison of Three Methods for Selecting Values of Input Variables in the Analysis of Output From a Computer Code, Technometrics 42 (2000) 55–61. doi:10.1080/00401706.2000.10485979

    arXiv 2000

  15. [15]

    Papachristodoulou, R

    L. Papachristodoulou, R. Martinez-Botas, T. Cao, A Meanline Loss Calibration Framework for the Design of Centrifugal Compressors, in: ASME Turbo Expo 2026: Turbomachinery Technical Conference and Exposition, American Society of Mechanical Engineers, Milan, Italy, 2026, pp. 1–16

    2026

  16. [16]

    Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, B. Poole, Score-Based Generative Modeling through Stochastic Differential Equations, 2021. doi:10.48550/arXiv.2011.13456.arXiv:2011.13456

    Pith/arXiv arXiv doi:10.48550/arxiv.2011.13456 2021

  17. [17]

    Ronneberger, P

    O. Ronneberger, P. Fischer, T. Brox, U-Net: Convolutional Networks for Biomedical Image Segmentation, 2015. doi:10.48550/arXiv. 1505.04597.arXiv:1505.04597

    Pith/arXiv arXiv doi:10.48550/arxiv.1505.04597 2015

  18. [18]

    Z. Liu, Y. Feng, M. J. Black, D. Nowrouzezahrai, L. Paull, W. Liu, MeshDiffusion: Score-based Generative 3D Mesh Modeling, 2023. doi:10.48550/arXiv.2303.08133.arXiv:2303.08133

    Pith/arXiv arXiv doi:10.48550/arxiv.2303.08133 2023

  19. [19]

    K. He, X. Zhang, S. Ren, J. Sun, Deep Residual Learning for Image Recognition, 2015. doi:10.48550/arXiv.1512.03385. arXiv:1512.03385

    Pith/arXiv arXiv doi:10.48550/arxiv.1512.03385 2015

  20. [20]

    H. G. Barrow, J. M. Tenenbaum, R. C. Bolles, H. C. Wolf, Parametric correspondence and chamfer matching: Two new techniques for image matching, in: Proceedings of the 5th International Joint Conference on Artificial Intelligence - Volume 2, IJCAI’77, Morgan Kaufmann Publishers Inc., San Francisco, CA, USA, 1977, pp. 659–663

    1977

  21. [21]

    Z. Wang, A. Bovik, H. Sheikh, E. Simoncelli, Image quality assessment: From error visibility to structural similarity, IEEE Transactions on Image Processing 13 (2004) 600–612. doi:10.1109/TIP.2003.819861

    arXiv 2004

  22. [22]

    M. L. Menéndez, J. A. Pardo, L. Pardo, M. C. Pardo, The Jensen-Shannon divergence, Journal of the Franklin Institute 334 (1997) 307–318. doi:10.1016/S0016-0032(96)00063-4

    doi:10.1016/s0016-0032(96)00063-4 1997

  23. [23]

    Y. Li, L. Liu, Enhancing Diffusion-based Point Cloud Generation with Smoothness Constraint, 2024. doi:10.48550/arXiv.2404.02396. arXiv:2404.02396

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2404.02396 2024

  24. [24]

    Juangphanich, Bio-Inspired Design of a Turbine Stage, Thesis, Purdue University, 2019

    P. Juangphanich, Bio-Inspired Design of a Turbine Stage, Thesis, Purdue University, 2019. doi:10.25394/PGS.10055423.v1

    work page doi:10.25394/pgs.10055423.v1 2019

  25. [25]

    F. R. Menter, Two-equation eddy-viscosity turbulence models for engineering applications, AIAA Journal 32 (1994) 1598–1605

    1994

  26. [26]

    S. Luo, W. Hu, Diffusion Probabilistic Models for 3D Point Cloud Generation, in: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021, pp. 2836–2844. doi:10.1109/CVPR46437.2021.00286

    arXiv 2021

  27. [27]

    X. Ren, J. Huang, X. Zeng, K. Museth, S. Fidler, F. Williams, XCube: Large-Scale 3D Generative Modeling using Sparse Voxel Hierarchies, in: 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 4209–4219. doi:10.1109/CVPR52733. 2024.00403

    arXiv 2024

  28. [28]

    J. Park, Y. Lee, J. Kim, Multi-modal conditional diffusion model using signed distance functions for metal-organic frameworks generation, Nature Communications 16 (2025) 34. doi:10.1038/s41467-024-55390-9. Y. Geng et al.:Preprint submitted to ElsevierPage 21 of 21

    doi:10.1038/s41467-024-55390-9 2025

This paper was first reviewed by grok-4.5 on July 30, 2026.