Parametric Shape Optimization

• Parametric Shape Optimization (PSO) is a method that defines optimal geometries using a finite set of design variables and controllable parameterization techniques.
• It leverages representations like NURBS, Bézier, and Fourier methods to balance geometric flexibility, smoothness, and computational tractability.
• PSO integrates optimization formulations, constraint management, and computational strategies—including gradient, evolutionary, and surrogate-based techniques—to solve real-world design challenges.

Parametric Shape Optimization (PSO) is the algorithmic process of determining the optimal geometry of an object or system—typically defined by a finite set of explicit (parametric) variables—to extremize a performance metric, often under constraints. The parameterization exploits a reduced, yet expressive, variable set (such as spline or shape-function coefficients, mesh node displacements, or design control points), mapping vectors in ℝⁿ directly to feasible shapes, with the aim of ensuring both geometric flexibility and computational tractability. PSO subsumes a wide class of engineering, physics, and mathematics applications, including aerodynamic/hydrodynamic drag minimization, thermal/metamaterial concentration, and the geometric optimization of biological morphologies.

1. Mathematical Foundations and Parameterization Approaches

The core of PSO is the finite-dimensional parameterization of a shape space. For a physical domain Ω, the mapping

$x(\xi) = \sum_{i=1}^n P_i N_i(\xi)$

using control points $P_i$ and basis functions $N_i$ (e.g., NURBS, Bézier, Legendre, or radial basis) defines the boundary or volume geometry (Jansari et al., 2022, Zhang et al., 2024, Ammad et al., 16 Jan 2026, Mallik et al., 2023). The choice of representation governs the smoothness, regularity, and admissible deformations:

• NURBS/Isogeometric (IGA): Delivers exact CAD-quality boundaries with controlled smoothness via fixed knot vectors and weights; control point displacements become the design variables (Jansari et al., 2022).
• Bézier and Generalized Trigonometric Bézier: Encode shapes by a few global control points, supporting high-fidelity surface interpolation (Zhang et al., 2024, Ammad et al., 16 Jan 2026).
• Fourier/Spherical Harmonics Support Functions: Suitable for convex and constant-width shapes, encoding global features with physically interpretable coefficients (Antunes et al., 2018).
• Level-Set and RBF Representations: Implicit boundary techniques, supporting both topology-preserving and topology-changing optimization; parameters may control global polynomial or local RBF terms (Mallik et al., 2023, Murea et al., 2020).
• FFD (Free-Form Deformation): Surface or volumetric morphing via lattice-based control point movement; extensible to complex 3D applications (Demo et al., 2021, Palazzolo et al., 2024, Demo et al., 2018, Demo et al., 2019).

Overall, parameterization selection establishes a trade-off between geometric expressivity, smoothness, computational expense, and compatibility with downstream numerical solvers.

2. Optimization Formulations and Constraints

A canonical PSO problem is formulated as

$$\begin{aligned} &\min_{\mathbf{x} \in \mathbb{R}^{N_{var}}} \quad f_{obj}(\mathbf{x}) \ &\text{subject to:} \quad h_i(\mathbf{x}) = 0; \quad g_j(\mathbf{x}) \le 0; \quad \mathbf{x}_{min} \leq \mathbf{x} \leq \mathbf{x}_{max} \end{aligned}$$

where $\mathbf{x}$ is the stacked vector of parametric variables (e.g., control-point coordinates), and $f_{obj}$ is often a functional derived from the solution of a PDE or other physical model (e.g., heat flux, total drag, resistance, or structural performance) (Jansari et al., 2022, Demo et al., 2021, Zhang et al., 2024, Mallik et al., 2023, Palazzolo et al., 2024, Alami et al., 1 Apr 2025).

Constraint classes include:

• Geometric: Self-intersection avoidance, symmetry, area/volume preservation, convexity, constant width, diameter bounds, and inclusion relations (Antunes et al., 2018, Jansari et al., 2022).
• Physical/Manufacturing: Stress/thickness constraints, minimum cure or offset thickness, fabrication bounds, or regulatory criteria (Altenhofen et al., 2019, Alami et al., 1 Apr 2025).
• Side Constraints: Trajectory straightness, non-collision, head-volume preservation for swimmer optimization (Palazzolo et al., 2024).

Constraint management may involve direct inclusion (hard constraints), penalization, projection, or explicit surrogate modeling.

3. Algorithmic and Computational Techniques

PSO encompasses a range of optimization algorithms, broadly classified into deterministic gradient-based solvers, derivative-free evolutionary methods, Bayesian optimization, and more recently, machine-learning–driven strategies.

• Particle Swarm Optimization (PSO): Implements position-velocity updates:

$$\begin{aligned} v_k^{i+1} &= \omega v_k^i + \alpha_1 \mathrm{rand}_1 (pbest_k^i - x_k^i) + \alpha_2 \mathrm{rand}_2 (gbest^i - x_k^i) \ x_k^{i+1} &= x_k^i + v_k^{i+1} \end{aligned}$$

with adaptive inertia and swarm-wide best tracking (Jansari et al., 2022, Ammad et al., 16 Jan 2026). Velocity and position clamps, convergence criteria, and box constraint enforcement are standard.

• Genetic Algorithms (GA), Active Subspace GAs: Population-based optimization, with embedding projection onto active directions (principal eigen-directions of the gradient covariance) to accelerate search in high-dimension spaces (Demo et al., 2021).
• Bayesian Optimization (BO): Utilizes Gaussian Process surrogates for both objectives and constraints, querying expected improvement acquisition functions and iteratively refining the design landscape via trust-region sampling (Palazzolo et al., 2024).
• LLM-based Evolutionary Strategies: LLMs (e.g., Claude 3.5 Sonnet) ingest few-shot records of prior best designs and propose new mean vectors for sampling, operating as a learned, zero-shot evolutionary operator (Zhang et al., 2024).
• Surrogate Modeling and Dimensionality Reduction: Reduced-Order Models (ROM), Proper Orthogonal Decomposition (POD), Gaussian Process Regression (GPR), and physics-informed parametric embeddings minimize the number of expensive PDE solves required (Demo et al., 2021, Demo et al., 2019, Serani et al., 2022, Serani et al., 8 Apr 2025, Mallik et al., 2023).
• Adjoint-Based and Sobolev-Smoothing Approaches: For differentiable cases, adjoint sensitivities feed reduced SQP and Sobolev-preconditioned Newton-like updates, leveraging elliptic Hessian approximations for larger, stable steps, particularly in One-Shot analysis-design cycles (Dick et al., 2021).

4. Coupling with Physical Models and Numerical Solvers

PSO inherently links geometric variables to physical field solutions. Popular couplings include:

• Isogeometric Analysis (IGA): NURBS bases parameterize both geometry and field (e.g., temperature), enabling direct use in high-order finite-element solvers (Jansari et al., 2022).
• CFD/FEM/FSI/Multiphysics: Navier-Stokes (steady/unsteady), RANS, Stokes flow, or Helmholtz equations are routinely solved on deformed geometries; mesh deformation is handled by FFD, RBF, or direct nodal updates (Demo et al., 2021, Demo et al., 2018, Alami et al., 1 Apr 2025, Demo et al., 2019, Jansari et al., 2022).
• Level-Set Hamiltonian Flow: Implicit approaches allow boundary and topology changes via evolution of the level-set function, with boundary integrals and sensitivities directly computed by tracing ODEs in parameter space (Murea et al., 2020).
• Surrogate Flow and Physics Prediction: Encoder-decoder CNNs are trained to map implicit parametrizations (e.g., level-set functions) to physical quantities orders of magnitude faster than traditional solvers, albeit with potential bias (Mallik et al., 2023, Serani et al., 8 Apr 2025).
• Multifidelity Optimization: Physics-informed or physics-driven embeddings align design variables with both geometry and low-fidelity physical predictions, enabling rapid screening and reduced search spaces (Serani et al., 8 Apr 2025, Serani et al., 2022).

5. Applications and Exemplary Benchmark Studies

The PSO framework is utilized across a spectrum of scientific disciplines:

• Thermal Metamaterial Design: NURBS-IGA+PSO approaches for maximizing mid-section heat flux in concentrator and cloak devices, achieving smooth, optimal CAD-like boundaries and outperforming large-scale FEM-level-set methods (Jansari et al., 2022).
• Aerodynamics and Hydrodynamics: Bézier control points, Legendre expansions, and FFD in airfoil and hull optimization, demonstrating drag reduction, lift–drag trade-off identification, and efficient navigation of multi-modal surfaces (Zhang et al., 2024, Demo et al., 2021, Demo et al., 2019, Mallik et al., 2023, Demo et al., 2018).
• Convex/Constant-Width Geometry: Support function parametrization and Fourier/spherical harmonics enable direct enforcement of convexity, diameter, or width constraints for Laplacian eigenvalue optimization, minimal-volume rotors, and the Meissner conjecture (Antunes et al., 2018).
• Bio-inspired and Microbial Locomotion: FFD head parameterization and analytic flagellar models, jointly optimized by Bayesian BO, recover biologically relevant pusher/puller morphologies and quantify speed-efficiency Pareto fronts in microswimming (Palazzolo et al., 2024).
• Additive-Subtractive Manufacturing: Optimization of overbuild/offset thickness and structural infill for minimized machining with geometric, manufacturing, and physical constraints, using sparse grid surrogates (Altenhofen et al., 2019).
• Plateau-type Problems: Generalized trigonometric Bézier patches, interior Dirichlet energy minimization, and PSO-tuned shape parameters yield surfaces with reduced area/energy relative to polynomial and harmonic counterparts under identical boundary data (Ammad et al., 16 Jan 2026).
• Hull and Propeller Optimization: Modal-truncated SVD or PCA embeddings of deformations, enriched with distributed/lumped physics and used for physics-informed dimension reduction and multi-objective screening (Demo et al., 2019, Serani et al., 8 Apr 2025).

Empirical results routinely demonstrate order-of-magnitude speed-ups over direct simulation-based loops, with surrogate-augmented or reduced-order pipelines providing rapid design space exploration for complex engineering systems.

6. Dimensionality Reduction, Embedding, and Data-Driven PSO

The scaling of PSO to high-dimensional or expensive-to-evaluate models is driven by advances in data-driven reduction:

• Parametric Model Embedding (PME): Applies block-weighted PCA to concatenated geometric and parametric (and optionally physical) data matrices, producing a latent subspace for optimization that retains an explicit invertible mapping to the native CAD parameterization (Serani et al., 2022, Serani et al., 8 Apr 2025). Physics-informed (PI-PME) and physics-driven (PD-PME) variants embed distributed/lumped signals, further aligning the search space with the directions responsible for physical objective variability.
• Active Subspace Identification: Eigen-decomposition of the gradient covariance matrix identifies the “most active” directions, enabling optimization in lower-dimensional projections with demonstrable convergence improvements in hull-form and marine cases (Demo et al., 2021, Demo et al., 2019).
• ROM/Machine-Learning Surrogates: POD, GPR, CNN, and kriging interpolate physical fields and objectives in reduced latent spaces, achieving real-time or near-real-time optimization, with further potential for machine-learning–based ROMs calibrated for physical consistency (Demo et al., 2021, Mallik et al., 2023, Serani et al., 8 Apr 2025).
• Physics-informed and multifidelity workflows: Integrating low-fidelity surrogate models for field/lumped quantities with PME-based optimization accelerates the screening of candidates and supports adaptive refinement with high-fidelity solvers (Serani et al., 8 Apr 2025).

7. Challenges, Limitations, and Future Directions

Key practical challenges persist:

• Parameterization Limitations: Linear (PCA-type) reductions may fail for highly nonlinear shape–response manifolds, or when the geometry-physics mapping is strongly coupled (Serani et al., 2022, Serani et al., 8 Apr 2025).
• Scalability: Topology-changing optimization remains difficult within most explicit parameterizations; implicit level-set or hybrid representations partially address this (Murea et al., 2020).
• Constraint Management: Enforcing geometric non-intersection, complex manufacturing constraints, or multi-objective trade-offs can introduce substantial complexity, often mitigated by penalty methods or staged surrogate models (Altenhofen et al., 2019, Alami et al., 1 Apr 2025).
• Surrogate Bias and Robustness: Machine-learnt surrogates may exhibit nonnegligible bias (e.g., ~10% in drag), necessitating downstream verification with full-order models (Mallik et al., 2023).
• Integration into CAD/FEM/CFD Pipelines: Direct compatibility with legacy geometry and simulation environments is enhanced by methods embedding the original CAD parameter semantics, but care must be taken to preserve geometric feasibility and enforce metamodel bounds (Serani et al., 2022, Serani et al., 8 Apr 2025).

Research continues to build out non-linear embedding techniques (autoencoders), multi-fidelity approaches, hybrid adjoint–surrogate methods, and the integration of modern LLMs for implicit search-space navigation (Zhang et al., 2024, Serani et al., 8 Apr 2025). The convergence of PSO with data-driven and physics-informed methods is expected to yield further gains in speed, reliability, and automation across industrial and scientific applications.