Isogeometric analysis

Isogeometric analysis is a class of computational methods in which functions originating in computer-aided design are used to represent both geometric domains and approximate solutions of partial differential equations. The approach was introduced to reduce the representational gap between geometric design models and conventional finite element analysis, where a CAD surface or solid is commonly converted into a separate polynomial mesh before simulation.

The standard formulation employs non-uniform rational B-splines, usually abbreviated NURBS, although later formulations also use T-splines, subdivision surfaces, hierarchical splines, and related technologies. These bases preserve important geometric properties of CAD models while supporting a Galerkin method for numerical approximation. Isogeometric analysis therefore belongs to the broader family of spline-based discretization methods rather than constituting a single numerical algorithm.

Historical development

The term was introduced in 2005 by Thomas J. R. Hughes, John A. Cottrell, and Yuri Bazilevs. Their formulation treated CAD geometry as the starting point of analysis instead of reconstructing that geometry with a conventional finite element mesh. The same basis functions used to parameterize the domain were used to approximate displacement, velocity, pressure, temperature, or another unknown field.

Early work concentrated on NURBS because they were already established in industrial geometric modeling and could exactly represent common analytic shapes such as conic sections. Subsequent research examined spline spaces with local refinement, since tensor-product NURBS propagate newly inserted knots across an entire parametric direction. Michael A. Scott and colleagues developed analysis-suitable T-splines that retained linear independence under controlled local refinement. John A. Evans and collaborators later established compatible spline complexes for differential operators appearing in incompressible flow and electromagnetic formulations.

The treatment of thin structures became another major area of development. Josef Kiendl and collaborators formulated isogeometric Kirchhoff–Love shell theory using the higher continuity available within spline patches. This avoided introducing independent rotational variables in formulations whose geometry and displacement fields possessed sufficient smoothness.

In 2016, You Watanabe and collaborators developed an isogeometric hydroelastic formulation for thin marine structures represented by trimmed NURBS surfaces. Their work coupled a spline-based shell discretization to a free-surface fluid model while retaining the parameterization of the structural design model. The formulation treated trim curves as integration boundaries and imposed fluid–structure interface conditions through a variational coupling operator. Its numerical studies addressed the influence of patch interfaces and geometric trimming on computed structural response.

Geometric representation

A univariate B-spline basis is defined by a nondecreasing knot vector

[ \Xi={\xi_1,\xi_2,\ldots,\xi_{n+p+1}}, ]

where (p) is the polynomial degree and (n) is the number of basis functions. The degree-zero functions are piecewise constant over knot spans, while functions of higher degree are generated recursively. For degree (p), the basis functions are

[ N_{i,p}(\xi)

\frac{\xi-\xi_i}{\xi_{i+p}-\xi_i}N_{i,p-1}(\xi) + \frac{\xi_{i+p+1}-\xi}{\xi_{i+p+1}-\xi_{i+1}}N_{i+1,p-1}(\xi), ]

with terms containing zero denominators interpreted as zero.

NURBS basis functions introduce positive weights (w_i) through the rational expression

[ R_i(\xi)

\frac{N_{i,p}(\xi)w_i} {\sum_j N_{j,p}(\xi)w_j}. ]

A curve is then represented as

[ \mathbf{x}(\xi)=\sum_i R_i(\xi)\mathbf{P}_i, ]

where (\mathbf{P}_i) are control points. Unlike nodal coordinates in an interpolatory finite element mesh, most control points do not lie on the represented curve. Their influence is distributed according to the support of the associated basis functions.

Surfaces and volumes are generally formed through tensor products. A surface parameterization has the form

[ \mathbf{x}(\xi,\eta)

\sum_{i=1}^{n} \sum_{j=1}^{m} R_{ij}(\xi,\eta)\mathbf{P}_{ij}, ]

where (R_{ij}) incorporates the univariate bases and their corresponding weights. The parametric domain is partitioned into knot spans, which play a role analogous to finite elements for quadrature and assembly. They do not, however, determine the complete support of a basis function, since a degree-(p) B-spline ordinarily extends across (p+1) knot spans.

Within a patch, an unrepeated interior knot generally gives continuity of class (C^{p-1}). Repeating a knot lowers continuity according to its multiplicity. This relation permits geometric features and solution irregularities to be represented by altering the spline space without changing the polynomial degree.

Variational discretization

Let (\Omega) be a physical domain obtained from a parametric domain (\widehat{\Omega}) through a spline mapping

[ \mathbf{x}=\mathbf{F}(\boldsymbol{\xi}). ]

For a scalar boundary-value problem, the approximate field is written as

[ u_h(\mathbf{x})

\sum_{A=1}^{N} R_A!\left(\mathbf{F}^{-1}(\mathbf{x})\right)d_A, ]

where (d_A) are algebraic degrees of freedom. The same collection of geometric basis functions may be used for (u_h), although enriched or mixed spaces can differ from the geometry space.

For the Poisson equation,

[ -\nabla\cdot(\kappa\nabla u)=f \quad\text{in }\Omega, ]

the Galerkin discretization produces the linear system

[ \mathbf{K}\mathbf{d}=\mathbf{f}, ]

with entries

[ K_{AB}

\int_{\Omega} \kappa,\nabla R_A\cdot\nabla R_B,d\Omega, \qquad f_A

\int_{\Omega} R_A f,d\Omega. ]

Numerical integration is normally conducted over knot spans after transforming each integral to parametric coordinates. The Jacobian of the geometric map enters both the integration measure and the transformation of basis-function derivatives. Consequently, singular or highly distorted parameterizations can affect an isogeometric discretization even when the visible physical boundary remains geometrically correct.

The resulting matrix structure differs from that of a low-order finite element discretization. Spline basis functions usually possess wider support, producing more nonzero interactions per degree of freedom. Their higher continuity also changes approximation and spectral properties, particularly for operators involving higher derivatives.

Refinement and continuity

Isogeometric refinement distinguishes among changes to the knot vector, polynomial degree, and continuity. Knot insertion increases the number of basis functions while preserving the represented geometry and polynomial degree. This operation is commonly called (h)-refinement because it decreases characteristic knot-span dimensions.

Degree elevation raises the polynomial order while preserving the geometry. When degree elevation is followed by knot insertion without reproducing every new knot multiplicity, the resulting operation increases both degree and continuity. This process is known as (k)-refinement and differs from standard finite element (p)-refinement, which generally retains element-interface continuity.

Tensor-product refinement affects an entire row or column of knot spans. T-splines and hierarchical B-splines alter this structure by allowing refinement to terminate within the parameter domain. Truncated hierarchical B-splines additionally modify coarse-level functions whose support overlaps refined regions, preserving partition of unity while limiting the overlap between refinement levels.

Continuity across separate geometric patches is not automatic. Adjacent patches may share a physical boundary while possessing incompatible parameterizations or distinct control meshes. Coupling can be imposed through direct identification when the discretizations match, through Lagrange multipliers, or through variational techniques such as Nitsche's method. The selected coupling changes the algebraic system and the regularity represented across the interface.

Trimmed geometry

Industrial CAD models often represent a physical surface as a portion of a larger parametric surface bounded by trim curves. The underlying tensor-product basis remains defined outside the retained region, while numerical integration is restricted to the visible part of the surface. This produces cut knot spans near trim boundaries and complicates the construction of accurate quadrature rules.

Trimming also separates the topological boundary from the original tensor-product parameterization. Basis functions intersected by a trim curve can possess very small active supports, which may affect matrix conditioning. Stabilized formulations and basis-removal criteria have been developed to control these effects, although such modifications must preserve approximation properties and boundary representation.

A trimmed boundary is not automatically equivalent to a watertight multipatch model. Two trimmed surfaces can coincide geometrically while retaining unrelated parameter coordinates. Analysis of the assembled model therefore requires an explicit treatment of continuity, interface integration, or both.

Structural and fluid formulations

Spline continuity is directly relevant to theories containing second-order spatial derivatives of the primary unknown. In Kirchhoff–Love shells, the weak form involves surface curvature derived from second derivatives of the displacement field. A sufficiently smooth spline basis can represent these derivatives without introducing rotational degrees of freedom within each regular patch.

In computational fluid dynamics, isogeometric spaces have been applied to incompressible Navier–Stokes equations. Equal-order velocity and pressure spaces do not generally satisfy the Ladyzhenskaya–Babuška–Brezzi condition, so stable formulations employ compatible spline spaces or add an appropriate stabilization term. Yuri Bazilevs and collaborators also integrated isogeometric discretization with variational multiscale methods for turbulent and moving-domain flows.

Fluid–structure interaction combines these considerations with the transfer of tractions and velocities across a moving interface. When the fluid and structural meshes are nonmatching, the coupling operator must account for the distinct trace spaces. A common geometric representation can simplify the description of the interface, but it does not by itself establish discrete conservation or numerical stability.

Numerical characteristics

The exact preservation of a CAD geometry applies when the analysis uses the original geometric representation without approximation or incompatible conversion. It does not imply that the computed physical field is exact. Discretization error remains determined by the solution space, quadrature, boundary conditions, and regularity of the governing problem.

Higher continuity can increase accuracy per degree of freedom for sufficiently smooth solutions, particularly in eigenvalue problems and higher-order differential equations. The same continuity broadens basis support and increases matrix bandwidth relative to many low-order finite element spaces. Computational cost therefore depends on the balance between the number of degrees of freedom, quadrature work, sparse matrix structure, and solver behavior.

Spline parameterizations can also contain geometric irregularities that are largely invisible in the physical model. Uneven knot spacing, extreme weights, and nearly singular mappings influence derivative transformations and conditioning. The quality of the parametric map is consequently part of the numerical model rather than only a matter of geometric presentation.

Relationship to finite element analysis

Isogeometric analysis retains the variational foundations of the finite element method. Both approaches construct finite-dimensional trial and test spaces, evaluate elementwise or spanwise integrals, assemble global algebraic systems, and impose boundary conditions in discrete form. Their principal distinction concerns the basis functions and the relationship between the computational space and the geometric model.

Conventional finite elements usually employ piecewise polynomial bases associated with mesh nodes or element-local degrees of freedom. Isogeometric methods instead use spline control variables and basis functions whose support may cross several knot spans. The boundary between the two categories is not absolute, since spline elements can be expressed through Bézier extraction and incorporated into element-based assembly frameworks.

Bézier extraction rewrites a spline basis over each knot span as a linear transformation of Bernstein polynomials. This representation separates local element evaluation from the global continuity encoded by the extraction operator. It provides an algebraic connection between spline spaces and traditional finite element software without changing the underlying approximation space.