PaperlayerView exact revision on arXiv

Source-linked AI summary

Aerodynamic Shape Design Space Exploration with Deep Latent Diffusion Model

Zhen Wei, Edouard Dufour, Colin Pelletier, Michaël Bauerheim, Pascal Fua

arXiv:2609.00812v1cs.CE

TL;DR

Aerodynamic design-space exploration must operate with scarce data while handling complex geometry and constraints. DiffGeo combines automatic latent-space shape parameterization with latent diffusion and energy-based conditioning, and reports data-efficient, reusable generation across airfoil and turbomachinery-blade cases. The framework is presented as a tool for refining and exploring existing turbomachinery design workflows, with that role forming an explicit scope boundary.

  • Problem

    Aerodynamic design-space exploration needs generation methods that remain useful under scarce data while addressing geometry validity, task adaptation, and complex constraints.

  • Method

    DiffGeo learns a latent shape representation, performs diffusion sampling in that space, and applies energy-based guidance for task-specific objectives and constraints.

  • Results

    Across airfoil and turbomachinery-blade studies, DiffGeo achieves high-quality, diverse, and controllable generation with at least an order of magnitude less data than alternatives.

  • Takeaways & Limitations

    DiffGeo decouples geometry representation from design targets, supporting reuse across tasks while incorporating complex constraints during sampling.

  • Takeaways & Limitations

    The study focuses on geometric design objectives, while extension to performance-based space exploration remains future work.

Abstract

from arXiv · show

We propose DiffGeo, a latent space diffusion-based generative framework for aerodynamic design space exploration under extreme data scarcity. DiffGeo combines a learned latent space model for automatic shape parameterization, with a diffusion sampler to directly generate novel, geometry-valid and controllable designs. We validate the approach on a series of case studies: (i) a 2D airfoil generation benchmark, where DiffGeo's latent diffusion model is compared against GAN- and VAE-based baselines in terms of sample quality, diversity and constraint adherence under limited data; (ii) integration into a surrogate-based optimization pipeline, where DiffGeo's conditional sampling produces task-informed airfoil data that improve both surrogate modeling and optimization performance; and (iii) extension to 3D turbomachinery blade prototyping, where DiffGeo generates realistic and high-performance blade geometries from a small set of reference designs. Throughout these investigations, DiffGeo achieves high-quality and diverse shape generation with at least an order of magnitude less data than alternatives, decouples geometry representation from design targets for flexible reuse, and seamlessly incorporates complex design constraints via energy-based conditioning. These capabilities demonstrate DiffGeo's potential to enhance early-stage design by automating design space exploration--improving efficiency, expanding design diversity and embedding engineering knowledge through controllable guidance.

Nomenclature

The paper frames DiffGeo as a latent-space diffusion framework for aerodynamic design-space exploration, targeting data scarcity, limited adaptability, and weak controllability in existing approaches.

  • Motivation: DiffGeo addresses data-driven DSE methods that require substantial data, adapt poorly to new objectives, and provide limited fine-grained control.Existing approaches also struggle with high-dimensional geometric variation and passive sampling toward feasible or high-performing regions.
  • Framework: DiffGeo combines an automatically learned shape latent space with diffusion-based sampling to generate valid aerodynamic geometries.The diffusion model operates on a shape latent space learned by an automatic parameterization method.
  • Framework: DiffGeo’s diffusion sampling is presented as more stable and data-efficient than GAN- and VAE-based approaches, without adversarial discriminators.The framework is intended for training in data-scarce environments.
  • Framework: DiffGeo separates geometry generation from task-specific objectives, allowing new differentiable energy functions to guide sampling without retraining the generator.This design supports reuse of the core shape sampler across different tasks.
  • Results: Experiments report successful training with as few as 50 airfoils or 75 three-dimensional blades, reducing data requirements by one to two orders of magnitude versus prior deep generative models.The reported examples use blades linearly interpolated from six base profiles.
  • Results: DiffGeo supports fine-grained conditioning with complex, vector-valued constraints, including constraints that existing generative design tools do not provide.The paper highlights multi-parameter control over generated geometry.

II. Related Work

Related work progresses from handcrafted parameterizations and iterative surrogate-assisted search toward learned latent representations and deep generative samplers, while retaining challenges in data efficiency, stability, and controllability.

  • Aerospace DSE: Classical aerodynamic DSE uses analytic parameter bases, DoE, and deformation methods, but these approaches impose structural assumptions or require crafted parameterizations and iterative loops.Surrogate-assisted methods reduce high-fidelity evaluations but remain dependent on parameterization and optimization design.
  • Aerospace DSE: Learned latent manifolds capture nonlinear variation among feasible shapes and restrict sampling toward valid geometries.This addresses the abnormal results often produced by random shape perturbations.
  • Deep Generative Design: Deep generative models directly synthesize novel designs, but their exploration depends on sampling strategy and surrogate-model quality.Earlier latent-space methods were paired primarily with static sampling or surrogate-based search.
  • Deep Generative Design: GAN-based aerodynamic generators can suffer from mode collapse, instability, substantial data demands, and architectural-tuning requirements.These limitations complicate reliable generation under data scarcity.
  • Deep Generative Design: Paired inverse-design models require curated geometry–performance datasets and carefully balanced losses, increasing data-engineering complexity.This constraint accompanies their ability to map target performance metrics to multiple geometries.
  • Diffusion Models: Recent conditional diffusion methods often require large labeled datasets, while lacking explicit parameterization can cause instability and post-processing needs.The cited limitations motivate DiffGeo’s latent-space and energy-guidance design.
  • Diffusion Models: DiffGeo positions itself as a reusable tool that automates and enhances existing MDO workflows rather than replacing them with a monolithic end-to-end model.Its stated role is rapid production of design candidates for downstream analysis and optimization.

C. Denoising Diffusion Models

DiffGeo performs diffusion in a compact latent manifold learned by an automatic shape parameterization model, then decodes sampled latent codes into regular, simulation-ready geometries.

  • Diffusion Background: Diffusion models learn reverse denoising from progressively corrupted data and can be formulated through SDEs, probability-flow ODEs, or continuous-time flows.The paper uses the latent-space formulation rather than describing a new diffusion formalism.
  • Latent Diffusion: Operating diffusion in compressed latent space can improve computational efficiency and training stability, motivating DiffGeo’s compact-manifold design.DiffGeo adopts latent diffusion rather than operating directly on high-dimensional geometry representations.
  • Latent Space Geometry Parameterization: The LSM learns an automatic parameterization that deforms a template mesh using a low-dimensional latent code.The deformed mesh is formed by adding predicted vertex displacements to template vertices while preserving edges.
  • Latent Space Geometry Parameterization: LSM training minimizes geometric discrepancy between reconstructed and target surfaces while regularizing latent codes for smoothness.Chamfer distance measures geometric difference, and the balancing weight controls latent-norm regularization.
  • Latent Space Geometry Parameterization: At inference, a target geometry is encoded by optimizing its latent vector with frozen LSM weights, then decoded into a mesh.The resulting mesh combines the template vertices, learned displacement, and template edges.
  • Latent Space Diffusion Model: Because the latent distribution is unknown and non-analytical, LSDM maps a feasible random distribution into valid latent vectors for decoding.This provides the generative sampling mechanism over the learned geometry manifold.

B. Latent Space Diffusion Model for Unconditional Generation

LSDM learns a diffusion process in the LSM latent space, denoising Gaussian noise into valid shape codes that decode into geometries. Conditional sampling adds differentiable energy guidance, with enhanced re-noising loops addressing imperfect objective satisfaction.

  • Unconditional generation: LSDM maps a standard multivariate normal distribution into the LSM latent space through a learned reverse denoising process.The model uses a forward noising process and learned reverse process represented as a Markov chain.
  • Unconditional generation: Unconditional sampling starts from z_T∼N (0, I), iteratively applies reverse transitions, and decodes z_0 into a shape.The reverse transitions use stochastic gradient Langevin dynamics.
  • Conditional generation: The conditional model incorporates equality and inequality constraints through a differentiable energy penalty evaluated on decoded geometries during reverse diffusion.Penalty coefficients weight equality and inequality violations, and the resulting energy gradient guides sampling toward lower-energy designs.
  • Conditional generation: Energy-based guidance augments the reverse-process score with a gradient that encourages geometries satisfying design objectives and constraints.The guidance strength is controlled by a temperature parameter ξ>0.
  • Enhanced conditional sampling: When complex objectives and tight constraints conflict with learned-distribution sampling within fixed diffusion steps, final samples may not fully satisfy the targets.The enhanced strategy is proposed specifically for this convergence limitation, especially in 3D shape generation.
  • Enhanced conditional sampling: Enhanced conditional sampling partially re-noises failed decoded geometries, restarts reverse diffusion, and repeats this loop to explore nearby alternative solution paths.The procedure uses an intermediate noise level T* and remains near the solution manifold during repeated regeneration.

D. Implementation Details

The implementation evaluates DiffGeo across data-limited 2D airfoil generation, surrogate-based optimization, and 3D blade prototyping, using shared latent geometry representations and several baselines.

  • Model configuration: The LSM latent dimension is d=256, selected to balance reconstruction capacity and computational efficiency.The LSM decoder is a 4-layer MLP with hidden widths [256, 256, 512, 512].
  • Model configuration: LSDM uses a 5-layer MLP with hidden widths [256, 512, 512, 512, 256] and T=1000 diffusion steps.The noise schedule increases β from 10−4 to 0.02.
  • Conditional configuration: Conditional sampling uses a time-dependent guidance schedule, with ξ=10^6β_t given as an example for area constraints.Enhanced conditional sampling uses N_T=3 loops and re-noising level T*=300.
  • Computational cost: With 50 airfoils, LSM and LSDM training take 671.9 seconds and 716.4 seconds, respectively, on an NVIDIA V100.Unconditional generation averages 690.51 milliseconds per 50 samples.
  • Experimental design: Three case studies examine data efficiency, deployment flexibility, and complex constraint handling in 2D and 3D design contexts.They include airfoil benchmarking, task-informed surrogate training data, and constrained 3D blade prototyping.
  • Experimental design: The experiments compare DiffGeo with GAN- and VAE-based models on sampling quality, novelty, diversity, and constraint adherence under limited data.All models use the same learned LSM representation to isolate sampling effects from geometry encoding.

1. Generative Adversarial Network Baseline

The baselines generate shapes in the shared LSM latent space using either adversarial sampling or learned Gaussian latent priors, enabling comparison with LSDM’s external diffusion sampler.

  • GAN baseline: The GAN baseline trains a generator–discriminator pair in the LSM latent space, then decodes generated latent samples into airfoils.The generator maps Gaussian noise to latent codes, while the discriminator distinguishes generated from training-derived codes.
  • GAN baseline: The conditional GAN accepts target parameters as conditioning inputs but requires paired data and retraining for new conditions.This differs from DiffGeo’s energy-based conditioning during sampling.
  • VAD baseline: The VAD baseline assigns each training shape a latent Gaussian distribution and samples new codes using the reparameterization trick.Its objective combines reconstruction through chamfer distance with KL regularization toward a standard normal prior.
  • Qualitative comparison: With fewer than 250 training samples, GAN outputs collapse toward nearly identical shapes, whereas VAD produces several similar-shape families.LSDM retains variation in thickness, camber, leading-edge radius, and maximum-thickness position even with 50 airfoils.
  • Qualitative comparison: LSDM latent codes overlap the real-airfoil distribution across training-set sizes, while GAN codes develop gaps and clusters associated with mode collapse.The contrast is especially pronounced below 250 training samples.

5. Quantitative Evaluation on Sampling Quality

Quantitative evaluations show that LSDM remains closest to the training distribution and maintains higher diversity, novelty, and geometric variability as training data decrease.

  • FID: LSDM achieves the lowest FID across all training-data sizes, indicating generated airfoils are statistically closest to the training-shape distribution.FID is computed in features extracted by a CFD-based surrogate model rather than an image classifier.
  • FID: With 50 airfoils, GAN’s FID is much higher, while VAD’s FID is higher than LSDM’s but more stable across dataset sizes.For more than 500 samples, GAN’s FID approaches LSDM’s, although LSDM retains a slight advantage.
  • Metric definitions: The diversity metric uses average chamfer distance to k=10 nearest generated neighbors, while novelty uses distance to k=10 nearest training examples.Higher novelty indicates generated shapes differ more from training designs while remaining valid.
  • Diversity and novelty: LSDM has the highest diversity and novelty scores across all evaluated scenarios and remains strong with only 50 training airfoils.GAN diversity and novelty collapse at 100 or 50 samples, while VAD remains lower than LSDM.
  • Geometric variability: LSDM and VAD maintain relatively flat geometric-variance curves as training data vary, whereas GAN variability drops sharply below 200 training shapes.GAN outputs then show nearly identical area, maximum thickness, and camber.

7. Discussion

DiffGeo’s latent diffusion sampler consistently outperforms GAN- and variational approaches in data-scarce airfoil generation, preserving diversity, novelty, fidelity, and constraint adherence. Conditional energy guidance maintains exploration while satisfying target geometry constraints.

  • DiffGeo’s LSDM outperforms GAN- and VAE-based approaches in sampling quality and exploration, especially under low-data settings.It offers stable generation without mode collapse and produces more diverse, higher-fidelity samples than prior-based alternatives.
  • Nearly an order of magnitude improvement in data efficiency over GANs is achieved for unconditional generation, with higher diversity and novelty than VAD.These advantages persist in the extreme low-data regime.
  • CLSDM has significantly lower area error than CGAN, indicating closer satisfaction of the target area even with very little training data.CGAN’s error is an order of magnitude higher and improves little with additional data.
  • Energy-based conditioning balances constraint adherence and output diversity, whereas CGAN often fails the condition or collapses sampling diversity.CLSDM’s novelty reflects exploration within the constrained design space rather than invalid off-target shapes.
  • DiffGeo decouples geometry generation from task-specific guidance, enabling conditional objectives without model retraining.The conditional benchmark uses energy-based guidance to enforce area during sampling without additional training.

B. Task-Informed Data Generation for Surrogate-Based Optimization

DiffGeo’s conditional sampling generates task-informed airfoil datasets from the same small historical dataset used by a baseline, improving surrogate accuracy and optimization outcomes. Its advantage comes from combining feasible, diverse shapes with task-specific thickness guidance.

  • Task-informed data generation: SM#3 trains on 50 historical airfoils and uses conditional DiffGeo sampling to produce 500 diverse, thickness-constrained shapes for surrogate training.The generated dataset embeds the thickness constraint while retaining exploration.
  • Surrogate performance: SM#3 reaches an R2 of 0.96 between surrogate-predicted and CFD-simulated L/D along the optimization path.This indicates near-perfect consistency with simulated performance throughout optimization.
  • Surrogate performance: SM#1-2 achieves R2 = 0.80, compared with R2 = 0.33 for SM#1-1 and R2 = 0.20 for SM#2.The results show that expert-tuned sampling outperforms wide-range and historical random sampling, although SM#3 retains a distinct advantage.
  • Optimization outcomes: SM#3 achieves the highest optimized L/D, while other strategies produce aggressive deformations associated with excessive drag and degraded actual performance.Its optimized airfoil remains a subtle modification of the baseline and follows a smoother pressure recovery.
  • Optimization outcomes: DiffGeo-generated data improve surrogate accuracy and guide optimization toward a superior design by generalizing around the baseline’s relevant design region.The reported workflow achieves R2 ≈0.96 and balances shape validity with sufficient exploration.
  • Scope boundary: The optimization study is limited to a finite local design set around NACA-0012 under inviscid Euler flow, so results validate the workflow rather than certify final transonic designs.The authors frame the optimized shapes as workflow-validation results within a simplified local region.

C. 3D Turbomachinery Blade Design Prototyping

DiffGeo extends to 3D turbomachinery blade prototyping using only 75 interpolated reference geometries. Its unconditional outputs are geometrically valid and statistically aligned with training blades, including their twist distributions.

  • Task setup: The 3D study evaluates whether DiffGeo can generate realistic, CFD-valid rotor blades from a small reference set while learning beyond a linear data space.The task also tests controllability under spanwise thickness and twist constraints.
  • Task setup: 75 training blades are constructed from convex combinations and binary merges of six baseline designs, producing physically plausible, meshable geometries.This dataset lies in a low-dimensional subspace and does not directly provide the original parametric representations.
  • Unconditional generation: Unconditionally generated blades exhibit smooth surfaces, feasible twist, and reasonable thickness and camber distributions without explicit geometric filtering.The outputs contain no non-physical artifacts.
  • Unconditional generation: DiffGeo-generated geometric statistics align closely with the 75-blade training distribution across thickness, camber, and section angles of attack.The comparison uses 200 generated blades against the 75 training blades.
  • Unconditional generation: Generated twist profiles vary individually, but their averaged twist law matches the mean twist law of the training blades.This supports statistical consistency of the learned 3D shape distribution.

3. Controllability of Guided Sampling

DiffGeo uses energy-based guidance to impose coupled 3D blade constraints while preserving unrelated geometric properties and diversity. It also generates blades outside the convex hull of the limited training set, demonstrating nonlinear design-space exploration.

  • Thickness-guided generation: DiffGeo imposes coupled hub and tip thickness constraints, including maximum-thickness and thickness-ratio limits, during guided sampling.The constraints are higher-dimensional and coupled, reflecting realistic turbomachinery design rules.
  • Thickness-guided generation: Guided blades remain smooth and diverse, while their thickness metrics shift toward the targets and camber and angle-of-attack metrics remain aligned with training data.This isolates the effect of thickness conditioning without distorting unrelated blade properties.
  • Nonlinear controllability: The generated residuals indicate that DiffGeo explores geometry outside the range accessible through convex combinations of the six base blades.The authors interpret nonzero residuals as evidence of nonlinear extrapolation rather than memorization or simple interpolation.
  • Nonlinear controllability: Unconditional generation has convex-hull residual 9.7, while thickness-guided generation reaches 24.9 average per-surface-point L2 distance.These correspond to roughly 12% and 31% of the typical hub chord, respectively.
  • Nonlinear controllability: DiffGeo achieves balanced, low total error across simultaneous in-convex-hull specifications, whereas linear interpolation struggles with the coupled targets.Linear interpolation also yields mean surface-point standard deviation 0.31, compared with 1.40 for DiffGeo under the same target.
  • Nonlinear controllability: DiffGeo’s nonlinear latent space provides additional degrees of freedom for design-space exploration while maintaining geometric validity and sample diversity.This advantage appears in the comparison with linear convex interpolation under the same targets.

5. Integrating DiffGeo with Mean-Line Design

DiffGeo is integrated with mean-line design to generate 3D rotor blades that satisfy prescribed twist while retaining variation in other geometric features. Conditional sampling and CFD evaluation indicate higher median efficiency than convex-hull interpolation, although the comparison is limited.

  • 3D blade workflow: DiffGeo automates candidate 3D blade generation from mean-line specifications, replacing manual construction with viable starting points for CFD analysis and optimization.The generated candidates satisfy prescribed twist requirements and can reduce design complexity and workload.
  • Mean-line guidance: Mean-line analysis supplies twist distributions and related aerodynamic quantities that are encoded as guidance during DiffGeo sampling.The twist law prescribes blade metal angle from hub to tip and supports reconstruction of quasi-3D sections.
  • Conditional sampling: 222 of 300 generated blades were valid and met the twist requirement, a 74% success rate after filtering invalid geometries.Strong twist guidance produced some less-smooth invalid shapes, but these were removed using visual or simple geometric checks.
  • CFD evaluation: 45 generated blades and 10 dataset blades met the pressure-ratio criterion 1.58 ≤ PR ≤ 1.6 for subsequent efficiency comparison.The selected candidates came from converged CFD simulations of generated and dataset blades.
  • Performance comparison: The generated blades achieved a median isentropic efficiency of 90.7%, compared with 89.3% for convex-hull interpolations.DiffGeo-generated blades consistently showed higher efficiency, and their upper efficiency range exceeded that of interpolated blades.
  • Caveat: The efficiency comparison is based on a limited number of selected candidates, so the reported gains should be interpreted as indicative.This scope boundary qualifies the comparison between conditional sampling and linear interpolation.
  • Nonlinear exploration: DiffGeo’s high-efficiency blades lie outside the base blades’ convex hull, indicating nonlinear latent-space variation rather than simple linear combinations.The mean reconstruction error was 5.9 measured as averaged per-surface-point L2 distance.

V. Conclusion

DiffGeo is presented as a data-efficient, controllable latent diffusion framework for aerodynamic design exploration and reusable integration into existing workflows. Across 2D airfoil and 3D blade studies, it generated valid, novel, constraint-satisfying designs and improved reported surrogate and aerodynamic outcomes, while remaining bounded by the learned data manifold and facing smoothness trade-offs under strong guidance.

  • Framework: DiffGeo uses latent-space diffusion with energy-based conditioning to generate diverse candidates that satisfy performance constraints in a one-shot, differentiable process.The framework is positioned as a reusable shape sampler for design space exploration.
  • Capabilities: The same pre-trained model can be repurposed for varying objectives without retraining, supporting task-agnostic adaptability.This adaptability is identified as a core capability in the 2D and 3D investigations.
  • Data efficiency: DiffGeo supports high-fidelity generation with as few as 50 training airfoil samples and 6 blade samples.The conclusion identifies this as evidence of high data efficiency under severe data limitations.
  • 2D airfoil optimization: In 2D surrogate-based optimization, guided generative sampling improved surrogate-model accuracy and final optimized performance over random or historical-data baselines.The reported comparison concerns baseline strategies based on random statistical sampling or historical data.
  • 3D blade study: In 3D blade studies, DiffGeo generated valid designs outside the original linear interpolation space and produced higher aerodynamic performance.The study used only six base profiles and demonstrated nonlinear extrapolation beyond their convex hull.
  • Limitations: The framework’s exploration is bounded by the training-data manifold, making it suited to interpolation and variation rather than invention of new topologies.This defines the paper’s scope boundary for conceptual design generation.
  • Limitations: Strong energy-based guidance creates a trade-off between strict constraint adherence and geometric smoothness.The paper also notes that physical design objectives have not yet been incorporated.
Loading 2609.00812v1…