Shape-Adaptive Reduced Basis Methods

• Shape-adaptive RBM is a model-order reduction technique that incorporates explicit geometry parameterization to efficiently simulate PDEs and boundary-integral problems under shape changes.
• It builds reduced spaces using methods like POD, greedy algorithms, and eigenfunction expansions, adapting the basis functions through parameter-to-geometry maps and Piola transforms.
• By separating offline heavy computations from rapid online evaluations, the approach significantly accelerates many-query simulations and optimization across various applications.

Searching arXiv for recent and foundational papers on shape-adaptive reduced basis methods and related geometric RBM formulations. {"query":"shape-adaptive reduced basis method geometry parameterized boundary integral reduced basis arXiv", "max_results": 10} {"query":"(Henríquez et al., 2024) Reduced Basis Method for the Elastic Scattering by Multiple Shape-Parametric Open Arcs in Two Dimensions", "max_results": 5} Searching by exact arXiv id and title to verify the primary source and gather related work. {"query":"id:(Henríquez et al., 2024)", "max_results": 5} Shape-adaptive reduced basis methods (RBMs) are model-order reduction procedures for parametrized PDE and boundary-integral problems in which geometric variation is part of the parameter space and the reduced approximation is constructed to remain effective under shape change. In the literature represented by Pinto–Henríquez’s treatment of elastic scattering by multiple open arcs, by reduced-basis optimization of plasmonic nanoparticle arrays, by rigorous source–mask optimization in 3D computational lithography, and by divergence-conforming reduced models for geometric Navier–Stokes flow, shape adaptivity is realized through parameter-to-geometry maps, reference-domain pullbacks, Piola transforms, or basis functions that depend explicitly on the current shape parameters (Henríquez et al., 2024, Gao et al., 6 Aug 2025, Pomplun et al., 2010, Fonn et al., 2018).

1. Geometric parametrization and the meaning of shape adaptivity

A central ingredient is an explicit parametrization of geometry. For multiple elastic cracks in two dimensions, one assumes $M$ disjoint open arcs $\Gamma_1,\dots,\Gamma_M$, with each arc $\Gamma_j$ given as the image of the reference interval $\hat S=[-1,1]$ under a smooth parameter-to-geometry map

$$X_j(\cdot;\mu_j):\hat S\to\mathbb R^2,\qquad s\mapsto X_j(s;\mu_j),$$

where $\mu_j\in\mathbb R^{P_j}$, and the full parameter is $$\mu=(\mu_1,\dots,\mu_M)\in\mathcal P\subset\mathbb R^P$$, $P=\sum_j P_j$. A common choice expands the shape around a nominal straight reference curve: $$X_j(s;\mu_j)=\xi_j(s)+\sum_{p=1}^{P_j}\mu_{j,p}\phi_{j,p}(s),\qquad s\in[-1,1].$$ The geometry-dependent normal $n_j(s;\mu_j)$ and metric factor $\Gamma_1,\dots,\Gamma_M$0 then enter the operators directly (Henríquez et al., 2024).

For plasmonic nanoparticle arrays, the configuration of $\Gamma_1,\dots,\Gamma_M$1 analytic nanoparticles is parameterized by

$\Gamma_1,\dots,\Gamma_M$2

where each $\Gamma_1,\dots,\Gamma_M$3 encodes the shape, for instance ellipse semi-axes, orientation, and center. In the 3D source–mask optimization setting, the parameter vector is split into geometry and illumination: $\Gamma_1,\dots,\Gamma_M$4 with $\Gamma_1,\dots,\Gamma_M$5 for mask geometry and $\Gamma_1,\dots,\Gamma_M$6 for the quadrupole source. In the incompressible-flow formulation, geometric dependence is expressed by a smooth invertible map $\Gamma_1,\dots,\Gamma_M$7, or equivalently $\Gamma_1,\dots,\Gamma_M$8, between a reference domain and the physical domain (Gao et al., 6 Aug 2025, Pomplun et al., 2010, Fonn et al., 2018).

Within these formulations, “shape-adaptive” does not denote a single algorithmic template. In some cases the reduced basis itself depends explicitly on the current geometry; in others the basis is fixed on a reference domain and transported to the physical domain by a geometry-dependent mapping. This suggests that shape adaptivity is best understood as a structural property of the reduced model rather than a specific discretization choice.

2. High-fidelity formulations for parametrized geometry

The full-order problems differ by application class, but each exposes geometry in a way that enables reduced modeling.

For elastic scattering by multiple open arcs, the exterior Lamé boundary-value problem is transformed by single-layer and double-layer potentials into a coupled system of boundary integral equations on $\Gamma_1,\dots,\Gamma_M$9: $\Gamma_j$0 For $\Gamma_j$1 on $\Gamma_j$2 and $\Gamma_j$3 on $\Gamma_j$4,

$\Gamma_j$5

$\Gamma_j$6

with analogous expressions for the hypersingular operator $\Gamma_j$7 and the adjoint $\Gamma_j$8. Here $\Gamma_j$9 is the elastic fundamental solution (Henríquez et al., 2024).

For plasmonic nanoparticles under the TM approximation, the forward problem is written as a parameterized layer-potential system for densities $\hat S=[-1,1]$0: $\hat S=[-1,1]$1 with parameterized single-layer and Neumann–Poincaré operators $\hat S=[-1,1]$2 and $\hat S=[-1,1]$3 (Gao et al., 6 Aug 2025).

In 3D computational lithography, the truth approximation is a finite-element discretization of time-harmonic Maxwell equations in a unit cell $\hat S=[-1,1]$4. Variationally, one seeks $\hat S=[-1,1]$5 such that

$\hat S=[-1,1]$6

where

$\hat S=[-1,1]$7

Discretization with edge elements yields the algebraic system

$\hat S=[-1,1]$8

In divergence-conforming Navier–Stokes reduction, the high-fidelity problem is first posed on the reference domain and discretized so that the velocity snapshots are pointwise divergence-free; the geometric dependence then enters the pulled-back forms and the Piola transport (Pomplun et al., 2010, Fonn et al., 2018).

A recurring theme is that the geometric parameters appear inside kernels, coefficients, domain maps, or test spaces. That feature is precisely what necessitates shape-aware reduction.

3. Construction of reduced spaces

The reduced spaces in shape-adaptive RBM are not uniform across applications.

In Pinto–Henríquez’s multiple-arc scattering framework, the offline stage begins by ignoring inter-arc coupling and solving $\hat S=[-1,1]$9 independent single-arc boundary-integral equations over a training set $$X_j(\cdot;\mu_j):\hat S\to\mathbb R^2,\qquad s\mapsto X_j(s;\mu_j),$$0. On arc $$X_j(\cdot;\mu_j):\hat S\to\mathbb R^2,\qquad s\mapsto X_j(s;\mu_j),$$1, the high-fidelity density is

$$X_j(\cdot;\mu_j):\hat S\to\mathbb R^2,\qquad s\mapsto X_j(s;\mu_j),$$2

Collecting all such solutions yields the snapshot set $$X_j(\cdot;\mu_j):\hat S\to\mathbb R^2,\qquad s\mapsto X_j(s;\mu_j),$$3. Proper Orthogonal Decomposition (POD) is then applied: with correlation matrix $$X_j(\cdot;\mu_j):\hat S\to\mathbb R^2,\qquad s\mapsto X_j(s;\mu_j),$$4 and dominant eigenpairs $$X_j(\cdot;\mu_j):\hat S\to\mathbb R^2,\qquad s\mapsto X_j(s;\mu_j),$$5, the POD basis functions are

$$X_j(\cdot;\mu_j):\hat S\to\mathbb R^2,\qquad s\mapsto X_j(s;\mu_j),$$6

These basis functions are orthonormal in $$X_j(\cdot;\mu_j):\hat S\to\mathbb R^2,\qquad s\mapsto X_j(s;\mu_j),$$7 and capture the largest energy of the snapshot set (Henríquez et al., 2024).

In source–mask optimization, the reduced space is built by a Greedy algorithm over a dense training set $$X_j(\cdot;\mu_j):\hat S\to\mathbb R^2,\qquad s\mapsto X_j(s;\mu_j),$$8. One initializes with a truth solution $$X_j(\cdot;\mu_j):\hat S\to\mathbb R^2,\qquad s\mapsto X_j(s;\mu_j),$$9, repeatedly computes reduced approximations $\mu_j\in\mathbb R^{P_j}$0 and estimators $\mu_j\in\mathbb R^{P_j}$1, selects

$\mu_j\in\mathbb R^{P_j}$2

and enlarges the basis by the newly computed truth solution. The resulting space is

$\mu_j\in\mathbb R^{P_j}$3

This is described as a self-adaptive RB construction for variable geometrical parameters (Pomplun et al., 2010).

The plasmonic absorber formulation develops a more explicit form of shape adaptation. Spectral theory for the Laplace NP-operator on each analytic particle $\mu_j\in\mathbb R^{P_j}$4 provides an orthogonal eigenbasis

$\mu_j\in\mathbb R^{P_j}$5

where $\mu_j\in\mathbb R^{P_j}$6 for an ellipse of semi-axes $\mu_j\in\mathbb R^{P_j}$7 and focal distance $\mu_j\in\mathbb R^{P_j}$8. Retaining the first $\mu_j\in\mathbb R^{P_j}$9 eigenmodes $$\mu=(\mu_1,\dots,\mu_M)\in\mathcal P\subset\mathbb R^P$$0, one approximates

$$\mu=(\mu_1,\dots,\mu_M)\in\mathcal P\subset\mathbb R^P$$1

Since the eigenfunctions depend explicitly on the shape parameters $$\mu=(\mu_1,\dots,\mu_M)\in\mathcal P\subset\mathbb R^P$$2, the resulting basis “adapts” as the optimizer moves in $$\mu=(\mu_1,\dots,\mu_M)\in\mathcal P\subset\mathbb R^P$$3 (Gao et al., 6 Aug 2025).

For steady Navier–Stokes flow, shape adaptivity is achieved differently. The reduced basis is formed from divergence-conforming high-fidelity velocity snapshots $$\mu=(\mu_1,\dots,\mu_M)\in\mathcal P\subset\mathbb R^P$$4 on the reference domain, followed by a POD in the $$\mu=(\mu_1,\dots,\mu_M)\in\mathcal P\subset\mathbb R^P$$5-seminorm with covariance matrix

$$\mu=(\mu_1,\dots,\mu_M)\in\mathcal P\subset\mathbb R^P$$6

The reduced basis functions are

$$\mu=(\mu_1,\dots,\mu_M)\in\mathcal P\subset\mathbb R^P$$7

and are transported to the physical domain by the Piola transform, which preserves divergence-freeness for every $$\mu=(\mu_1,\dots,\mu_M)\in\mathcal P\subset\mathbb R^P$$8 (Fonn et al., 2018).

4. Offline–online decomposition and affine compression

The practical value of RBM depends on separating expensive parameter-independent work from inexpensive parameter-dependent work.

For multiple open arcs, the principal difficulty is nonaffinity: the operators $$\mu=(\mu_1,\dots,\mu_M)\in\mathcal P\subset\mathbb R^P$$9 and $P=\sum_j P_j$0 depend on $P=\sum_j P_j$1 through the geometry inside the kernel. A modified Empirical Interpolation Method (EIM) is applied to each kernel

$P=\sum_j P_j$2

to obtain

$P=\sum_j P_j$3

Hence

$P=\sum_j P_j$4

where $P=\sum_j P_j$5 is a nominal parameter used to fix the metric in the basis operators. With $P=\sum_j P_j$6 denoting the matrix of POD basis vectors, the reduced system is

$P=\sum_j P_j$7

The offline complexity for assembling and compressing the $P=\sum_j P_j$8 basis operators is $P=\sum_j P_j$9, while the POD costs $$X_j(s;\mu_j)=\xi_j(s)+\sum_{p=1}^{P_j}\mu_{j,p}\phi_{j,p}(s),\qquad s\in[-1,1].$$0. Online, evaluating the coefficients costs $$X_j(s;\mu_j)=\xi_j(s)+\sum_{p=1}^{P_j}\mu_{j,p}\phi_{j,p}(s),\qquad s\in[-1,1].$$1, assembling $$X_j(s;\mu_j)=\xi_j(s)+\sum_{p=1}^{P_j}\mu_{j,p}\phi_{j,p}(s),\qquad s\in[-1,1].$$2 costs $$X_j(s;\mu_j)=\xi_j(s)+\sum_{p=1}^{P_j}\mu_{j,p}\phi_{j,p}(s),\qquad s\in[-1,1].$$3, and solving the reduced system by a direct method costs $$X_j(s;\mu_j)=\xi_j(s)+\sum_{p=1}^{P_j}\mu_{j,p}\phi_{j,p}(s),\qquad s\in[-1,1].$$4, giving an online cost of order $$X_j(s;\mu_j)=\xi_j(s)+\sum_{p=1}^{P_j}\mu_{j,p}\phi_{j,p}(s),\qquad s\in[-1,1].$$5 (Henríquez et al., 2024).

In the plasmonic setting, the offline–online split is organized around singularity extraction. The singular part $$X_j(s;\mu_j)=\xi_j(s)+\sum_{p=1}^{P_j}\mu_{j,p}\phi_{j,p}(s),\qquad s\in[-1,1].$$6 is wavelength-independent and assembled offline from NP-eigenfunction captures, while the smooth remainder $$X_j(s;\mu_j)=\xi_j(s)+\sum_{p=1}^{P_j}\mu_{j,p}\phi_{j,p}(s),\qquad s\in[-1,1].$$7 is assembled online by trapezoidal rule for each $$X_j(s;\mu_j)=\xi_j(s)+\sum_{p=1}^{P_j}\mu_{j,p}\phi_{j,p}(s),\qquad s\in[-1,1].$$8. The forward reduced system reads

$$X_j(s;\mu_j)=\xi_j(s)+\sum_{p=1}^{P_j}\mu_{j,p}\phi_{j,p}(s),\qquad s\in[-1,1].$$9

and the adjoint problem has the analogous decomposition

$n_j(s;\mu_j)$0

By expanding in NP-eigenfunctions, the $n_j(s;\mu_j)$1 singularity is isolated in closed form (Gao et al., 6 Aug 2025).

In the Maxwell and Navier–Stokes formulations, the standard RBM affine decomposition is used. For lithography,

$n_j(s;\mu_j)$2

so that the reduced matrices $n_j(s;\mu_j)$3 and vectors $n_j(s;\mu_j)$4 are precomputed offline, and the online assembly of

$n_j(s;\mu_j)$5

costs $n_j(s;\mu_j)$6, followed by an $n_j(s;\mu_j)$7 solve. For geometric incompressible flow, the pulled-back forms admit affine expansions such as

$n_j(s;\mu_j)$8

and the online cost is $n_j(s;\mu_j)$9, independent of the high-fidelity dimension (Pomplun et al., 2010, Fonn et al., 2018).

5. Certification, stability, and methodological distinctions

A defining property of RBM in these works is the presence of computable error surrogates or preserved structural constraints.

For elastic scattering by shape-parametric open arcs, the reduced residual is

$\Gamma_1,\dots,\Gamma_M$00

and the certified error bound is

$\Gamma_1,\dots,\Gamma_M$01

where $\Gamma_1,\dots,\Gamma_M$02 is a lower bound for the inf–sup constant of the full operator and can be certified offline by a Successive Constraint Method. In the 3D Maxwell setting, the residual is

$\Gamma_1,\dots,\Gamma_M$03

and if $\Gamma_1,\dots,\Gamma_M$04, then

$\Gamma_1,\dots,\Gamma_M$05

In practice one uses a cheaply computable lower bound $\Gamma_1,\dots,\Gamma_M$06, and the estimator drives the Greedy basis generation (Henríquez et al., 2024, Pomplun et al., 2010).

The plasmonic absorber paper defines residual-based estimators

$\Gamma_1,\dots,\Gamma_M$07

and states that standard RBM bounds of the form

$\Gamma_1,\dots,\Gamma_M$08

provide certification. In the Navier–Stokes setting, the main structural issue is not an a posteriori bound but stability under geometric variation: the Piola transform guarantees that $\Gamma_1,\dots,\Gamma_M$09 for all $\Gamma_1,\dots,\Gamma_M$10, and pressure is recovered a posteriori at negligible extra cost (Gao et al., 6 Aug 2025, Fonn et al., 2018).

Several common misconceptions are clarified by these constructions. Shape adaptivity does not require repeated remeshing of a full-order model: in the nanoparticle optimization framework, no remeshing or velocity-field perturbation is required because shape sensitivities enter only through derivatives of the boundary operators. Nor does geometry parametrization necessarily destroy incompressibility: in the divergence-conforming approach, exact solenoidality is retained under geometric transformations by the Piola transform. Similarly, in the multi-arc scattering problem, the offline stage need not resolve the full coupled configuration; the basis is built from decoupled single-arc problems, while the inter-arc interaction is represented later through affine compression (Gao et al., 6 Aug 2025, Fonn et al., 2018, Henríquez et al., 2024).

6. Applications and reported numerical behavior

The reported numerical results show that shape-adaptive RBM is used both for rapid many-query simulation and for optimization loops with geometric parameters.

For elastic scattering by multiple open arcs, representative tests with $\Gamma_1,\dots,\Gamma_M$11 arcs and $\Gamma_1,\dots,\Gamma_M$12 shape parameters per arc show that the maximum reduced-basis error

$\Gamma_1,\dots,\Gamma_M$13

decays exponentially in $\Gamma_1,\dots,\Gamma_M$14, reaching $\Gamma_1,\dots,\Gamma_M$15 accuracy by $\Gamma_1,\dots,\Gamma_M$16. With $\Gamma_1,\dots,\Gamma_M$17, $\Gamma_1,\dots,\Gamma_M$18, $\Gamma_1,\dots,\Gamma_M$19, and $\Gamma_1,\dots,\Gamma_M$20, the online solve takes $\Gamma_1,\dots,\Gamma_M$21, versus $\Gamma_1,\dots,\Gamma_M$22 for a full BIE. As $\Gamma_1,\dots,\Gamma_M$23 increases, the RB approach scales nearly linearly in $\Gamma_1,\dots,\Gamma_M$24, while the full BIE scales like $\Gamma_1,\dots,\Gamma_M$25; the bound $\Gamma_1,\dots,\Gamma_M$26 remains tight even as $\Gamma_1,\dots,\Gamma_M$27 increases (Henríquez et al., 2024).

For broadband absorber design with plasmonic nanoparticles, the online stage with $\Gamma_1,\dots,\Gamma_M$28 and $\Gamma_1,\dots,\Gamma_M$29 is reduced by two orders of magnitude relative to classical BEM. In single- and multiple-particle tests, the RBM achieves relative errors below $\Gamma_1,\dots,\Gamma_M$30 in extinction cross-sections and adjoint fields, whereas a 200-point Nyström reference gives $\Gamma_1,\dots,\Gamma_M$31–$\Gamma_1,\dots,\Gamma_M$32. Online solutions of both forward and adjoint problems are $\Gamma_1,\dots,\Gamma_M$33 faster than classical BEM at comparable accuracy, and the projected gradient-descent loop converges in $\Gamma_1,\dots,\Gamma_M$34 iterations for broadband targets over $\Gamma_1,\dots,\Gamma_M$35 nm (Gao et al., 6 Aug 2025).

In 3D source–mask optimization, the truth problem has $\Gamma_1,\dots,\Gamma_M$36 degrees of freedom and one full solution takes approximately $\Gamma_1,\dots,\Gamma_M$37 s. A reduced basis with $\Gamma_1,\dots,\Gamma_M$38 snapshots yields $\Gamma_1,\dots,\Gamma_M$39-norm error $\Gamma_1,\dots,\Gamma_M$40 and diffraction-order output error below $\Gamma_1,\dots,\Gamma_M$41 at $\Gamma_1,\dots,\Gamma_M$42, with an online cost of approximately $\Gamma_1,\dots,\Gamma_M$43 s per solve, corresponding to a speed-up of about $\Gamma_1,\dots,\Gamma_M$44. An SMO run with approximately $\Gamma_1,\dots,\Gamma_M$45 forward solves is reduced from approximately $\Gamma_1,\dots,\Gamma_M$46 s to approximately $\Gamma_1,\dots,\Gamma_M$47 s, and the optimal parameters enlarge the process window by about $\Gamma_1,\dots,\Gamma_M$48 compared with mask-only optimization (Pomplun et al., 2010).

For steady flow around a NACA0015 airfoil, using $\Gamma_1,\dots,\Gamma_M$49 divergence-free velocity modes yields relative velocity error $\Gamma_1,\dots,\Gamma_M$50 and relative pressure error $\Gamma_1,\dots,\Gamma_M$51, uniformly over sampled angles of attack. The divergence-conforming Piola-based RB runs in roughly $\Gamma_1,\dots,\Gamma_M$52 s per query, versus on the order of $\Gamma_1,\dots,\Gamma_M$53 s for a Taylor–Hood supremizer-stabilized RB of comparable accuracy, and remains exactly divergence-free to machine precision for every $\Gamma_1,\dots,\Gamma_M$54 (Fonn et al., 2018).

Taken together, these results indicate that shape-adaptive RBM is not confined to a single physics domain. It appears in boundary-integral scattering, nanoparticle inverse design, rigorous electromagnetic lithography, and incompressible flow, with the common objective of replacing expensive geometry-dependent solves by low-dimensional, rapidly evaluable, and in several cases certified surrogates.