betaFeedback

WorksComputer Science

Immersed boundary-conformal isogeometric methods for magnetostatics

Abstract

Isogeometric analysis was proposed to bridge the gap between computer-aided design and numerical discretization. However, standard multi-patch isogeometric analysis mandates conformal discretizations across patch interfaces, posing challenges for multi-material domain problems. In the context of electric machines, this requirement becomes evident in a large number of patches needed to represent machines consisting of several domains and materials. In this work, we adopt, extend, and evaluate three non-conformal discretization strategies for magnetostatic problems: a fully immersed approach, the union with non-conformal patches, and the union with conformal layers. In all three methods, boundary-conformal high-order quadrature rules are employed for integration over trimmed boundary and interface elements. In the two union approaches, material regions are, as far as possible, represented by independent non-conformal spline patches that are embedded within a background patch and coupled weakly through Nitsche's method. In the latter framework, critical interfaces are additionally surrounded by conformal layers that enable the strong imposition of boundary conditions and improved resolution of interface behavior. The proposed approaches are assessed through several magnetostatic benchmark problems and an industrial application. The numerical results show that the union methods achieve highly accurate solutions, while the fully immersed approach struggles with discontinuities in field gradients across material interfaces. Nevertheless, these methods significantly reduce the geometric preprocessing effort compared to conventional, conformal multi-patch analysis and require substantially fewer patches. Overall, this demonstrates that our immersed boundary-conformal isogeometric framework possesses great potential for efficient simulation of complex electromagnetic devices.

Graphical Abstract

Image not extracted

Figure 18.

Highlights

Immersed boundary-conformal isogeometric methods for magnetostatics

Yusuf T. Elbadry, Giuliano Guarino, Pablo Antolín, Oliver Weeger

  • Accurate integration over trimmed elements via boundary-conformal quadrature method

  • Reduced geometry preprocessing effort through minimization of the required patch count

  • Mitigated need for Lagrange multiplier-type methods for coupling

  • Accurate solution for complex geometries without imposing conformality conditions

  • Demonstrated efficacy of the proposed method for magnetostatic applications

Introduction

Numerical simulation has become an essential tool in engineering and applied sciences, reducing overall design time, and enabling detailed analysis across a wide range of applications, such as solid mechanics [1, 2, 3, 4], fluid dynamics [5, 6, 7], and electric machines [8, 9, 10]. However, geometry preprocessing still constitutes a significant portion of the overall simulation process. The finite element method (FEM) represents one of the cornerstones of numerical simulation and is widely used across a broad range of engineering applications. Nonetheless, it suffers from limitations related to the approximation of geometric representations, particularly in the presence of curved boundaries and complex features. To address this issue, isogeometric analysis (IGA) was introduced to establish a tighter integration between computer-aided design (CAD) and finite element simulations [11, 12], by employing Non-Uniform Rational B-Splines (NURBS) for the simultaneous representation of both geometry and solution fields [13]. Despite these advances, geometric preprocessing remains a time-consuming step, especially for complex domains. Generation of body-fitted finite element meshes typically requires substantial manual intervention. In the context of IGA, conformal multi-patch discretizations must be employed [14]. This limits the degree of automation achievable in the overall simulation workflow, also in view of upcoming agent-based and large language model-empowered computer-aided engineering [15].

To mitigate the meshing effort, immersed boundary methods (IBM) provide an alternative approach to body-fitted, conformal meshes. A large number of methods adopt this concept, including the level-set method [16, 17], marker-and-cell methods [18, 19], volume-of-fluid methods [20, 21], cut-cell methods [22], fictitious domain and penalty methods [23, 24], the eXtended finite element method [25, 26, 27, 28], CutFEM [29], and the finite cell method [30]. However, the main challenge with such approaches lies in accurately resolving the geometry, particularly the interfaces between domains, either with adaptive discretizations or just the integration scheme, which are both also challenging. Moreover, Dirichlet boundary conditions must typically be imposed weakly while maintaining stability of the formulation.

In the context of magnetostatics and electric machine analysis, simulations typically involve multiple computational domains composed of different materials with strong material property contrasts, e.g., of the magnetic permeability of air vs. iron. Consequently, the meshing process becomes increasingly challenging, as material interfaces must be explicitly identified and accurately resolved. When multi-patch discretizations are employed, this results in a large number of patches, as the patches must be topologically square and the spline discretizations conformal across interfaces [14, 31, 32]. Constructing these patches, enforcing conformity conditions, and assembling them into a multi-patch mesh constitutes a tedious and time-consuming process, which may require several hours or even days of manual labor to complete for geometries with complex features. Alternatively, to loosen at least the conformity requirement, Nitsche’s method [33] can be used to couple multi-patches in a weak manner [34]. However, this can be applied to domains topologically equivalent to a square, a condition that is not met in complex geometries. Hence, such a technique eliminates the conformity condition, but large number of patches will be needed to represent a complex geometry. Furthermore, when rotational motion is considered, as in rotor–stator configurations, the coupling of interfaces between rotating and stationary subdomains becomes particularly challenging, since maintaining mesh conformity typically necessitates repeated re-meshing of the rotating regions [35, 10].

Multiple methods have addressed the coupling of rotating subdomains effectively by weakly coupling the interface between the rotor-stator subdomains. Mortar-type methods [36, 37] were applied to Maxwell’s equations, including non-overlapping mortar formulations [38, 35, 39]. For overlapping subdomains, mortar-based techniques have been applied to coupled magneto-mechanical problems [40]. A local discontinuous Galerkin method was proposed to treat the air-gap region in electric machine simulations [41]. In addition, interface treatment based on Lagrange multipliers has been investigated in [42, 43, 44], adopting non-conformal mapping approaches in [45, 46]. Beyond Lagrange multiplier-based techniques, an implicit boundary finite element method was proposed in [47] for magnetostatic analysis; in this method, material interfaces are handled without explicitly conforming meshes. In the context of IGA for electric machines, harmonic rotor-stator coupling techniques [48] have been combined with mortar methods [10]. In these approaches, the rotor and stator are constructed independently using conforming multi-patch discretizations that are subsequently coupled weakly through harmonic basis functions. However, while these methods effectively address the interface coupling between rotating and stationary domains, they generally still require mesh conformity within each subdomain (e.g., within the rotor and stator regions). This limits their applicability to complex geometries and reduces the overall level of automation.

To overcome challenges related to geometric pre-processing, patch conformity, and the accurate resolution of interfaces between geometric patches, the concept of the immersed boundary conformal method (IBCM) was introduced [49, 50]. The method was applied to linear elasticity [49], contact problems [51], and linear shell theory [50]. In this concept, multiple non-conforming patches are coupled weakly by means of Nitsche’s method, thereby avoiding the need for fully matching discretizations at interfaces. For integration over trimmed elements and interfaces, a high-order quadrature rule is employed. The approach can be further extended by surrounding the non-conformal patches with conformal boundary layers, which enables the strong imposition of Dirichlet boundary conditions, and smooths the transition between different material domains.

In the present work, we transfer three different concepts to two-dimensional magnetostatics, namely: fully immersed isogeometric analysis [52], the union with non-conformal patches (“Union NC”), and the union with conformal layers (“Union CL”) [49, 50]. We aim to establish a computational framework, in which complex geometries involving multiple materials can be represented using a minimal number of patches, without enforcing discretization conformity between them. In this way, strict multi-patch and conformal meshing requirements are circumvented while maintaining solution accuracy. The performance of the proposed approaches is assessed through a series of benchmark problems that feature, in contrast to previous works on heat diffusion or linear elasticity [49, 50], multiple material subdomains with challenges such as discontinuous material coefficients, abrupt variations in current density across subdomain interfaces, and a remanence source term that involves the gradient of the test functions.

The remainder of this manuscript is organized as follows: Section 2 introduces the governing equations of magnetostatics as foundational model, the bilinear form, and the usage of Nitsche’s method for weakly imposition of essential boundary conditions and multi-domain coupling. A brief overview of isogeometric finite element analysis is presented in Section 3, in addition to the discretized weak form. Section 4 then describes the three different numerical techniques, namely: fully immersed, union with non-conformal patches, and union with conformal layers. Moreover, a note is given on the topic of stability and conditioning. In Section 5, the proposed techniques are numerically evaluated through a series of benchmark examples, including configurations with impressed currents, material discontinuities, and a complex, industry-relevant multi-material geometry. Finally, Section 6 concludes the paper and outlines directions for future research.

Fundamentals of magnetostatics

Basics of magnetostatics

The mathematical modeling of electric machines is governed by Maxwell’s equations. However, in the case of low-frequency rotating machines, particularly when displacement currents and induced eddy currents can be neglected, the full set of Maxwell’s equations can be reduced to the magnetostatic formulation [53, 54].

We consider a physical domain Ω⊂R3\Omega\subset\mathbb{R}^{3}, on which the magnetostatic formulation is given by:

∇×H=J,(1)\nabla\times\mathbf{H} = \mathbf{J}, \tag*{(1)}
∇⋅B=0,(2)\nabla\cdot\mathbf{B} = 0, \tag*{(2)}

where H\mathbf{H} denotes the magnetic field strength, B\mathbf{B} is the magnetic flux density, and J\mathbf{J} represents the source current density. In this work, we assume all materials to be isotropic and linear. Thus, the field strength H\mathbf{H} and the flux density B\mathbf{B} are linked through the constitutive equation:

B=μH,(3)\mathbf{B} = \mu\mathbf{H}, \tag*{(3)}

where μ\mu is the material’s magnetic permeability. It is independent of the field strength and given as μ=μrμ0\mu= \mu_r \mu_0, with μr\mu_r denoting the relative permeability of the material and μ0=4π⋅10−7 H/m\mu_0 = 4\pi\cdot10^{-7}\ \mathrm{H}/\mathrm{m} being the permeability of air.

For materials exhibiting remanent behavior, such as permanent magnets, the constitutive relation reads [55]:

B=μH+Br,(4)\mathbf{B} = \mu\mathbf{H} + \mathbf{B}_r, \tag*{(4)}

where Br\mathbf{B}_r denotes the permanent magnet’s remanence.

To explicitly satisfy the divergence-free property of the magnetic flux density B\mathbf{B}, we introduce the vector potential function A\mathbf{A} such that:

B=∇×A.(5)\mathbf{B} = \nabla\times\mathbf{A}. \tag*{(5)}

Plugging (5) and the constitutive relation (4) into eq:1 yields:

∇×(ν(∇×A−Br))=J,(6)\nabla\times\left(\nu\left(\nabla\times\mathbf{A} - \mathbf{B}_r\right)\right) = \mathbf{J}, \tag*{(6)}

where ν=μ−1\nu= \mu^{-1} is the magnetic reluctivity.

As commonly done for modeling rotating electric machines with uniform cross-sections, we restrict our analysis to two-dimensional geometries Ω⊂R2\Omega\subset\mathbb{R}^{2} [56, 53]. Accordingly, for the two-dimensional formulation, only the out-of-plane, zz-components of the magnetic vector potential AzA_z and the current density JzJ_z are considered. Under these assumptions, the strong form of the magnetostatic problem reads:

−∇⋅(ν(∇Az−Br⊥))=Jzin Ω,(7)-\nabla\cdot\left(\nu\left(\nabla A_z-\mathbf{B}_{r}^{\perp}\right)\right)=J_z \quad\text{in }\Omega, \tag*{(7)}

where the in-plane remanence Br⊥\mathbf{B}_{r}^{\perp} is parameterized as Br⊥=Br(−sin⁡θr,cos⁡θr)⊤\mathbf{B}_{r}^{\perp}=B_r(-\sin\theta_r,\cos\theta_r)^{\top} with BrB_r being the remanence magnitude and θr\theta_r the remanence direction. To complete the boundary value problem, the following boundary conditions are imposed:

Az=gon ΓD⊂∂Ω,(8)A_z=g \quad\text{on }\Gamma_D\subset\partial\Omega, \tag*{(8)}
n⋅(∇Az)=0on ΓN⊂∂Ω,(9)\mathbf{n}\cdot(\nabla A_z)=0 \quad\text{on }\Gamma_N\subset\partial\Omega, \tag*{(9)}

where n\mathbf{n} is the outward unit normal to the boundary and ΓD∪ΓN=∂Ω\Gamma_D\cup\Gamma_N=\partial\Omega with ΓD∩ΓN=∅\Gamma_D\cap\Gamma_N=\varnothing.

Weak form of 2D magnetostatics

Applying the standard procedure of multiplying the strong form of the partial differential equation (7) with a test function vv, integrating over the domain Ω\Omega, performing integration by parts, applying the divergence theorem, and exploiting the boundary conditions given in Eq. (7), one obtains the weak form of the 2D magnetostatic problem: Find Az∈Hg1A_z\in H_g^1 such that

∫Ων∇v⋅∇Az dΩ=∫ΩvJz dΩ+∫Ων∇v⋅Br⊥ dΩ∀v∈H01.(10)\int_{\Omega}\nu\nabla v\cdot\nabla A_z\,\mathrm{d}\Omega=\int_{\Omega}vJ_z\,\mathrm{d}\Omega+\int_{\Omega}\nu\nabla v\cdot\mathbf{B}_{r}^{\perp}\,\mathrm{d}\Omega\quad\forall v\in H_0^1. \tag*{(10)}

Here, the boundary integral resulting from the integration by parts vanishes due to the imposition of the homogeneous Neumann condition on ΓN\Gamma_N, which physically represents a boundary where the magnetic flux is directed purely in the normal direction. Furthermore, the trial and test function spaces are defined as:

Hg1={u∈H1(Ω):u=g on ΓD},(11)H_g^1=\{u\in H^1(\Omega):u=g\ \text{on }\Gamma_D\}, \tag*{(11)}
H01={v∈H1(Ω):v=0 on ΓD}.(12)H_0^1=\{v\in H^1(\Omega):v=0\ \text{on }\Gamma_D\}. \tag*{(12)}

Using from now on u=Azu=A_z, the weak form in Eq. (10) can be written in a more concise, abstract mathematical form as: Find u∈Hg1u\in H_g^1 such that

a(u,v)=l(v)∀v∈H01,(13)a(u,v)=l(v)\quad\forall v\in H_0^1, \tag*{(13)}

with the symmetric positive definite bilinear form:

a(u,v)=∫Ων∇v⋅∇u dΩ,(14)a(u,v)=\int_{\Omega}\nu\nabla v\cdot\nabla u\,\mathrm{d}\Omega, \tag*{(14)}

and the linear form:

l(v)=∫ΩvJz dΩ+∫Ων∇v⋅Br⊥ dΩ.(15)l(v)=\int_{\Omega}vJ_z\,\mathrm{d}\Omega+\int_{\Omega}\nu\nabla v\cdot\mathbf{B}_{r}^{\perp}\,\mathrm{d}\Omega. \tag*{(15)}

Note here that, as linear elliptic partial differential equations, the bilinear forms of heat diffusion, linear elasticity, and magnetostatics are similar, but magnetostatics poses the additional challenges that (i) usually subdomains with different material properties must be modeled, resulting in discontinuously varying reluctivity ν\nu, source current density JzJ_z, and remanence Br⊥\mathbf{B}_{r}^{\perp} (which are often non-zero) and (ii) the remanence term in the linear form includes the gradient of the test function ∇v\nabla v.

Weak imposition of Dirichlet boundary conditions

In immersed boundary method settings, coincidence of the physical domain boundaries with the computational domain boundaries is not guaranteed. In such cases, strong imposition of the Dirichlet boundary conditions is not possible. There are different methods for weak imposition of the Dirichlet boundary conditions, such as the penalty method [57], Nitsche’s method [33], and Lagrange multiplier methods [58]. In this work, in case that strong imposition of Dirichlet boundary conditions u=gu = g on ΓD\Gamma_{D} is not possible, we apply Nitsche’s method for preserving the consistency and symmetry of the weak form.

Consequently, the bilinear form and linear form given in Eqs. (13) to (15) are amended to incorporate the additional penalty, consistency, and symmetry terms: Find u∈H1u \in H^{1} such that

a(u,v)+ag(u,v)=l(v)+lg(v)∀v∈H1,(16)a(u,v) + a_{g}(u,v) = l(v) + l_{g}(v) \quad\forall v \in H^{1}, \tag*{(16)}

where

ag(u,v)=−∫ΓDν(∇u⋅n)v dΓ−∫ΓDν(∇v⋅n)u dΓ+β∫ΓDνvu dΓ,(17)a_{g}(u,v) = -\int_{\Gamma_{D}} \nu(\nabla u \cdot\mathbf{n})v\,\mathrm{d}\Gamma- \int_{\Gamma_{D}} \nu(\nabla v \cdot\mathbf{n})u\,\mathrm{d}\Gamma+ \beta\int_{\Gamma_{D}} \nu vu\,\mathrm{d}\Gamma, \tag*{(17)}

and

lg(v)=−∫ΓDν(∇v⋅n)g dΓ+β∫ΓDνvg dΓ.(18)l_{g}(v) = -\int_{\Gamma_{D}} \nu(\nabla v \cdot\mathbf{n})g\,\mathrm{d}\Gamma+ \beta\int_{\Gamma_{D}} \nu vg\,\mathrm{d}\Gamma. \tag*{(18)}

Here, β>0\beta> 0 represents the stabilization parameter, which needs to be selected sufficiently large to impose the prescribed Dirichlet boundary condition.

Weak coupling of non-conformal domains

For coupling non-conformal domains along their interface, e.g., in the setting of non-conformal multipatch discretizations, we also employ Nitsche’s method [33], leveraging its consistent and symmetric formulation to weakly enforce interface continuity conditions.

For illustrating the concept in the two-dimensional magnetostatic setting, we consider the coupling of two disjunct domains ΩL\Omega^{L} and ΩR\Omega^{R}, ΩL∩ΩR=∅\Omega^{L} \cap\Omega^{R} = \varnothing, across their interface ΓLR=∂ΩL∩∂ΩR\Gamma_{LR} = \partial\Omega^{L} \cap\partial\Omega^{R}, see Figure 1. Though discontinuities in material parameters, magnetic remanences or source current densities can arise at the interfaces, the normal component of the magnetic flux density B\mathbf{B} and the tangential component of the magnetic field strength H\mathbf{H} should be continuous [59 Ch. I5]:

Coupling of two domains $\Omega^{L}$ and $\Omega^{R}$ across their common interface $\Gamma_{LR}$

Figure 1. Coupling of two domains ΩL\Omega^{L} and ΩR\Omega^{R} across their common interface ΓLR\Gamma_{LR} via Nitsche’s method

[[B]]⋅n=0⇒n×[[Az]]=0on ΓLR,(19)[[\mathbf{B}]] \cdot\mathbf{n} = 0 \quad\Rightarrow\quad\mathbf{n} \times[[A_{z}]] = 0 \quad\text{on } \Gamma_{LR}, \tag*{(19)}
n×[[H]]=0⇒n×[[ν∇×Az]]=0on ΓLR,(20)\mathbf{n} \times[[\mathbf{H}]] = 0 \quad\Rightarrow\quad\mathbf{n} \times[[\nu\nabla\times A_{z}]] = 0 \quad\text{on } \Gamma_{LR}, \tag*{(20)}

with interface jumps on ΓLR\Gamma_{LR} defined as [[∙]]=∙R−∙L[[\bullet]] = \bullet^{R} - \bullet^{L}. The outward unit normal to the interface ΓLR\Gamma_{LR} is defined as n=nR=−nL\mathbf{n} = \mathbf{n}^{R} = -\mathbf{n}^{L}, where nL\mathbf{n}^{L} and nR\mathbf{n}^{R} are the outward unit normals to ∂ΩL\partial\Omega^{L} and ∂ΩR\partial\Omega^{R}, respectively. In terms of the magnetic scalar potential u=Azu = A_{z}, these conditions translate as:

[[u]]=uL−uR=0on ΓLR,(21)[[u]] = u^{L} - u^{R} = 0 \quad\text{on } \Gamma_{LR}, \tag*{(21)}
[[ν∇u]]⋅n=νL∇uL⋅n+νR∇uR⋅n=0on ΓLR,(22)[[\nu\nabla u]] \cdot\mathbf{n} = \nu^{L}\nabla u^{L} \cdot\mathbf{n} + \nu^{R}\nabla u^{R} \cdot\mathbf{n} = 0 \quad\text{on } \Gamma_{LR}, \tag*{(22)}

where uR=u∣ΩRu^{R}=u|_{\Omega^{R}} and uL=u∣ΩLu^{L}=u|_{\Omega^{L}} denote the potentials in their corresponding domains, and νL\nu^{L} and νR\nu^{R} are the magnetic reluctivities in ΩL\Omega^{L} and ΩR\Omega^{R}, respectively.

To weakly enforce these two conditions using Nitsche’s method, the bilinear form in Eq. (13) is adapted: Find u∈Vgu \in V^{g} such that:

a(u,v)+an(u,v)=l(v),∀v∈V0.(23)a(u,v)+a_n(u,v)=l(v), \qquad\forall v \in V^{0}. \tag*{(23)}

Here, the global trial and test function space are defined as:

Vα={v∈H1(ΩL):v∣ΓD∩∂ΩL=α}⊕{v∈H1(ΩR):v∣ΓD∩∂ΩR=α},(24)V^{\alpha}=\left\{v \in H^{1}(\Omega^{L}):v|_{\Gamma_{D}\cap\partial\Omega^{L}}=\alpha\right\}\oplus\left\{v \in H^{1}(\Omega^{R}):v|_{\Gamma_{D}\cap\partial\Omega^{R}}=\alpha\right\}, \tag*{(24)}

where α:ΓD→R\alpha:\Gamma_{D}\to\mathbb{R} denotes the prescribed Dirichlet boundary condition on the joint Dirichlet boundary of ∂ΩL\partial\Omega^{L} and ∂ΩR\partial\Omega^{R}, such that it is α=g\alpha=g for VgV^{g} and α=0\alpha=0 for V0V^{0}, respectively. More importantly, the additional bilinear form to weakly enforce Eq. (17) by Nitsche’s method is given as:

an(u,v)=−∫ΓLR{ν∇u⋅n}γ2[[v]] dΓ−γ1∫ΓLR{ν∇v⋅n}γ2[[u]] dΓ+β(hR−1+hL−1)∫ΓLRνmax⁡[[u]][[v]] dΓ,(25)\begin{aligned} a_n(u,v)={}&-\int_{\Gamma_{LR}}\left\{\nu\nabla u\cdot\mathbf{n}\right\}_{\gamma_2}[[v]]\,\mathrm{d}\Gamma-\gamma_1\int_{\Gamma_{LR}}\left\{\nu\nabla v\cdot\mathbf{n}\right\}_{\gamma_2}[[u]]\,\mathrm{d}\Gamma\\ &+\beta\left(h_R^{-1}+h_L^{-1}\right)\int_{\Gamma_{LR}}\nu_{\max}[[u]][[v]]\,\mathrm{d}\Gamma, \tag*{(25)} \end{aligned}

with:

{ν∇u⋅n}γ2:=γ2(νR∇uR⋅n)+(1−γ2)(νL∇uL⋅n).(26)\left\{\nu\nabla u\cdot\mathbf{n}\right\}_{\gamma_2}:=\gamma_2\left(\nu^{R}\nabla u^{R}\cdot\mathbf{n}\right)+(1-\gamma_2)\left(\nu^{L}\nabla u^{L}\cdot\mathbf{n}\right). \tag*{(26)}

This latter expression represents a weighted flux jump on the interface ΓLR\Gamma_{LR}, compare Eq. (17b). The parameter γ2∈[0,1]\gamma_2\in[0,1] controls the weighting of the average operator and determines toward which domain the average is biased. In the present formulation, if we assumed that ΩR\Omega^{R} is discretized in an untrimmed, boundary-conformal fashion, i.e., as an individual patch, whereas domain ΩL\Omega^{L} is immersed and thus trimmed, then the choice γ2=1\gamma_2=1 leads to a one-sided flux evaluation, thereby eliminating the symmetric averaging procedure. Consequently, the penalty term depends solely on the mesh size hRh_R of the untrimmed “master” domain ΩR\Omega^{R}, rather than on an averaged mesh parameter that also considers the mesh size hLh_L of the trimmed domain ΩL\Omega^{L}. In our formulation, it will be essential that the selected master domain remains untrimmed in order to preserve stability. More elaboration on the trimming process and the union approach will be discussed in Sections 4.1 and 4.2.

In Eq. (25), the parameter γ1∈{−1,0,1}\gamma_1\in\{-1,0,1\} controls the symmetry of the formulation. Here, we set γ1=1\gamma_1=1 to maintain symmetry of the bilinear form ana_n. Furthermore, the stabilization parameter β>0\beta>0 in the last term of Eq. (25) is required to ensure coercivity. It must be chosen sufficiently large to enforce the interface condition in a stable and consistent manner. Here, we select β∈[10,104]\beta\in[10,10^{4}]. Additionally, the reluctivity used in the stabilization term is chosen as the maximum value of the reluctivity of the two domains in order to ensure robustness of the formulation, νmax⁡=max⁡{νL,νR}\nu_{\max}=\max\{\nu^{L},\nu^{R}\}.

Isogeometric analysis

This section introduces the basic concepts of isogeometric finite element analysis and fixes the notation used throughout the manuscript. Further details can be found in [13, 11, 12]. First, we introduce B-splines, followed by an isogeometric finite element discretization of the magnetostatic weak form Eq. (10).

B-splines

Let {ξ1,…,ξn+p+1}\{\xi_1,\ldots,\xi_{n+p+1}\} be a knot vector defining the parametric domain Ωξ=[ξ1,ξn+p+1]\Omega_{\xi}=[\xi_1,\xi_{n+p+1}]. The Cox–de Boor recursion formula is used to define the nn B-spline basis functions Nip:Ωξ→RN_i^p:\Omega_{\xi}\to\mathbb{R} of degree p≥0p\geq0 as:

Ni0(ξ)={1,if ξi≤ξ<ξi+1,0,otherwise.(27)N_i^0(\xi)= \begin{cases} 1, & \text{if }\xi_i\leq\xi<\xi_{i+1},\\ 0, & \text{otherwise}. \end{cases} \tag*{(27)}
Nip(ξ)=ξ−ξiξi+p−ξiNip−1(ξ)+ξi+p+1−ξξi+p+1−ξi+1Ni+1p−1(ξ).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).

Here, it is assumed that any quotient of the form 0/00/0 is taken to be zero. The B-spline basis functions possess several important properties, including non-negativity, linear independence, and partition of unity. Moreover, their regularity can be controlled by the knot multiplicity; in particular, they are (p−1)(p-1)-times continuously differentiable at interior knots of multiplicity one. To enforce the interpolation property at the boundary, open knot vectors are employed, for which the first and last knots are repeated (p+1)(p+1)-times.

Bivariate B-splines basis functions Ni(ξ)N_i(\boldsymbol{\xi}) with ξ=(ξ,η)\boldsymbol{\xi}=(\xi,\eta) can be constructed by the tensor product of two univariate B-spline basis functions as

Ni(ξ,η)=Nj,p(ξ)⋅Nk,q(η)for i=(j−1)l+kwith{j=1,…,l,k=1,…,m(28)N_i(\xi,\eta)=N_{j,p}(\xi)\cdot N_{k,q}(\eta)\quad\text{for }i=(j-1)l+k\quad\text{with}\quad \begin{cases} j=1,\ldots,l,\\ k=1,\ldots,m \end{cases} \tag*{(28)}

Here, Nj,p:Ωξ→RN_{j,p}:\Omega_{\xi}\to\mathbb{R} and Nk,q:Ωη→RN_{k,q}:\Omega_{\eta}\to\mathbb{R} are the ll and mm univariate B-spline basis functions of degree pp and qq, respectively, as defined in Eq. (27). The n=l⋅mn=l\cdot m bivariate B-splines basis functions Ni(ξ):Ω^→RN_i(\boldsymbol{\xi}):\hat{\Omega}\to\mathbb{R} are then defined on the two-dimensional parameter domain Ω^=Ωξ×Ωη\hat{\Omega}=\Omega_{\xi}\times\Omega_{\eta}. For simplicity of the notation, i=1,…,ni=1,\ldots,n is used as a combined index and the dependence of NiN_i on the degrees pp and qq is dropped. In the following, for the sake of clarity, we always set p=qp=q.

Isogeometric discretization

A two-dimensional B-spline surface Ω⊂R2\Omega\subset\mathbb{R}^{2} is represented through the mapping F:Ω^→Ω\mathcal{F}:\hat{\Omega}\to\Omega as a linear combination of nn bivariate B-splines basis functions Ni(ξ)N_i(\boldsymbol{\xi}) with control points xi∈R2\mathbf{x}_i\in\mathbb{R}^{2}:

F(ξ)=xh(ξ)=∑i=1nNi(ξ)xi.(29)\mathcal{F}(\boldsymbol{\xi})=\mathbf{x}^{h}(\boldsymbol{\xi})=\sum_{i=1}^{n}N_i(\boldsymbol{\xi})\mathbf{x}_i. \tag*{(29)}

The B-splines surface parametrization is not limited to describe the geometry in terms of the position vector xh∈Ω\mathbf{x}^{h}\in\Omega, but also to discretize and approximate the field variable and test functions:

Az≈uh(ξ)=∑i=1nNi(ξ)ui,v≈vh(ξ)=∑i=1nNi(ξ)vi.(30)A_z\approx u^{h}(\boldsymbol{\xi})=\sum_{i=1}^{n}N_i(\boldsymbol{\xi})u_i,\qquad v\approx v^{h}(\boldsymbol{\xi})=\sum_{i=1}^{n}N_i(\boldsymbol{\xi})v_i. \tag*{(30)}

Here, ui,vi∈Ru_i,v_i\in\mathbb{R} are the nn control points of the discretized scalar potential field uh:Ω^→Ru^{h}:\hat{\Omega}\to\mathbb{R} and the scalar test function vh:Ω^→Rv^{h}:\hat{\Omega}\to\mathbb{R}, respectively.

Under the isoparametric assumption, the field variables on the physical domain Ω\Omega are obtained by composing the discrete field with the inverse of the geometric mapping Eq. (29), e.g., uh(x)=uh(ξ)∘F(ξ)−1u^{h}(\mathbf{x})=u^{h}(\boldsymbol{\xi})\circ\mathcal{F}(\boldsymbol{\xi})^{-1}, such that uh,vh∈H1(Ω)u^{h},v^{h}\in H^{1}(\Omega). Hence, the gradients with respect to the physical domain coordinates needs to be computed using the chain rule, e.g.,

∇xuh(ξ)=∇ξuh(ξ)⋅(∇ξF(ξ))−1,(31)\nabla_{\mathbf{x}}u^{h}(\boldsymbol{\xi})=\nabla_{\xi}u^{h}(\boldsymbol{\xi})\cdot\left(\nabla_{\xi}\mathcal{F}(\boldsymbol{\xi})\right)^{-1}, \tag*{(31)}

with

∇ξuh(ξ)=∑i=1n∇ξNi(ξ)ui,J(ξ)=∇ξF(ξ)=∑i=1n∇ξNi(ξ)xi,(32)\nabla_{\xi}u^{h}(\boldsymbol{\xi})=\sum_{i=1}^{n}\nabla_{\xi}N_i(\boldsymbol{\xi})u_i,\qquad\mathbf{J}(\boldsymbol{\xi})=\nabla_{\xi}\mathcal{F}(\boldsymbol{\xi})=\sum_{i=1}^{n}\nabla_{\xi}N_i(\boldsymbol{\xi})\mathbf{x}_i, \tag*{(32)}

where J(ξ)\mathbf{J}(\boldsymbol{\xi}) here refers to the Jacobian of the geometric parameterization (not to be confused with the source current density, which is also denoted as J\mathbf{J} in Eq. (1), but reduced to the use of JzJ_z for 2D magnetostatics, see Eqs. (7) and (10)).

Isogeometric finite elements

Now, the discretization introduced in Eq. (30) is substituted into the magnetostatic weak form given in Eq. (10), such that the bilinear and linear form can be written as:

a(uh,vh)=∫Ων ∇xvh⋅∇xuh dΩ=∫Ω^ν(∇ξvh⋅J−1)⋅(∇ξuh⋅J−1)det⁡J dΩ^=∑i=1nvi∑j=1nuj∫Ω^(∇ξNi⋅J−1)⋅(∇ξNj⋅J−1)det⁡J dΩ^,(33)\begin{aligned} a(u^{h},v^{h})&=\int_{\Omega}\nu\,\nabla_{\mathbf{x}}v^{h}\cdot\nabla_{\mathbf{x}}u^{h}\,\mathrm{d}\Omega =\int_{\hat{\Omega}}\nu\left(\nabla_{\xi}v^{h}\cdot\mathbf{J}^{-1}\right)\cdot\left(\nabla_{\xi}u^{h}\cdot\mathbf{J}^{-1}\right)\det\mathbf{J}\,\mathrm{d}\hat{\Omega}\\ &=\sum_{i=1}^{n}v_i\sum_{j=1}^{n}u_j\int_{\hat{\Omega}}\left(\nabla_{\xi}N_i\cdot\mathbf{J}^{-1}\right)\cdot\left(\nabla_{\xi}N_j\cdot\mathbf{J}^{-1}\right)\det\mathbf{J}\,\mathrm{d}\hat{\Omega}, \tag*{(33)} \end{aligned}
l(vh)=∫ΩvhJz dΩ+∫Ων∇vh⋅Br⊥ dΩ=∫Ω^vhJzdet⁡J dΩ^+∫Ω^ν(∇ξvh⋅J−1)⋅Br⊥det⁡J dΩ^=∑i=1nvi∫Ω^(NiJz+ν(∇ξNi⋅J−1)⋅Br⊥)det⁡J dΩ^.(34)\begin{aligned} l(v^h) &= \int_{\Omega} v^h J_z\,\mathrm{d}\Omega+ \int_{\Omega} \nu\nabla v^h \cdot\mathbf{B}^{\perp}_r\,\mathrm{d}\Omega= \int_{\widehat{\Omega}} v^h J_z \det\mathbf{J}\,\mathrm{d}\widehat{\Omega} + \int_{\widehat{\Omega}} \nu\left(\nabla_{\xi}v^h\cdot\mathbf{J}^{-1}\right)\cdot\mathbf{B}^{\perp}_r \det\mathbf{J}\,\mathrm{d}\widehat{\Omega}\\ &= \sum_{i=1}^{n}v_i\int_{\widehat{\Omega}}\left(N_iJ_z+\nu\left(\nabla_{\xi}N_i\cdot\mathbf{J}^{-1}\right)\cdot\mathbf{B}^{\perp}_r\right)\det\mathbf{J}\,\mathrm{d}\widehat{\Omega}. \tag*{(34)} \end{aligned}

Here, the integration over the physical domain Ω\Omega and the evaluation of all shape functions and fields is pulled back to the parametric domain Ω^\widehat{\Omega}, as Ω=F(Ω^)\Omega=\mathcal{F}(\widehat{\Omega}).

Since a(uh,vh)=l(vh)a(u^h,v^h)=l(v^h) must hold for all vhv^h, the viv_i can be chosen arbitrarily, which results in the linear system of equations Ku=b\mathbf{K}\mathbf{u}=\mathbf{b}, where u=(u1,…,un)⊤∈Rn\mathbf{u}=(u_1,\ldots,u_n)^{\top}\in\mathbb{R}^n is the sought vector of the control points of the scalar magnetic potential and the coefficients of the system matrix K\mathbf{K} and the RHS vector b\mathbf{b} are given as:

Kij=∫Ω^ν(∇ξNi⋅J−1)⋅(∇ξNj⋅J−1)det⁡J dΩ^,(35)K_{ij}=\int_{\widehat{\Omega}}\nu\left(\nabla_{\xi}N_i\cdot\mathbf{J}^{-1}\right)\cdot\left(\nabla_{\xi}N_j\cdot\mathbf{J}^{-1}\right)\det\mathbf{J}\,\mathrm{d}\widehat{\Omega}, \tag*{(35)}
bi=∫Ω^(NiJz+ν(∇ξNi⋅J−1)⋅Br⊥)det⁡J dΩ^.(36)b_i=\int_{\widehat{\Omega}}\left(N_iJ_z+\nu\left(\nabla_{\xi}N_i\cdot\mathbf{J}^{-1}\right)\cdot\mathbf{B}^{\perp}_r\right)\det\mathbf{J}\,\mathrm{d}\widehat{\Omega}. \tag*{(36)}

Thanks to the piecewise polynomial definition and limited support of the B-spline shape functions NiN_i, for the numerical evaluation of the integrals arising from the discretized weak form, the parametric domain Ω^\widehat{\Omega} is partitioned into nen_e disjoint elements Ω^e⊂Ω^\widehat{\Omega}^e\subset\widehat{\Omega} such that ⋃e=1neΩ^e=Ω^\bigcup_{e=1}^{n_e}\widehat{\Omega}^e=\widehat{\Omega}. These elements are defined by the knot spans of the non-repeated knot vectors {ξ1,…,ξl+p+1}×{η1,…,ηm+q+1}\{\xi_1,\ldots,\xi_{l+p+1}\}\times\{\eta_1,\ldots,\eta_{m+q+1}\}, for example Ω^e=[ξj,ξj+1]×[ηk,ηk+1]\widehat{\Omega}^e=[\xi_j,\xi_{j+1}]\times[\eta_k,\eta_{k+1}].

The corresponding elements in the physical domain are given by Ωe=F(Ω^e)\Omega^e=\mathcal{F}(\widehat{\Omega}^e). Analogously to the isoparametric finite element method, a tensor-product Gauss-Legendre quadrature rule with nqp=(p+1)2n_{qp}=(p+1)^2 quadrature points is employed on each element [60, 11]. Ultimately, the numerical integration of Eqs. (35) and (36) leads to:

Kij≈∑e=1ne∑r=1nqpwre(∇ξNi(ξre)⋅J−1(ξre))⋅(∇ξNj(ξre)⋅J−1(ξre))det⁡J(ξre),(37)K_{ij}\approx\sum_{e=1}^{n_e}\sum_{r=1}^{n_{qp}}w_r^e\left(\nabla_{\xi}N_i(\xi_r^e)\cdot\mathbf{J}^{-1}(\xi_r^e)\right)\cdot\left(\nabla_{\xi}N_j(\xi_r^e)\cdot\mathbf{J}^{-1}(\xi_r^e)\right)\det\mathbf{J}(\xi_r^e), \tag*{(37)}
bi≈∑e=1ne∑r=1nqpwre(Ni(ξre)Jz(ξre)+ν(∇ξNi(ξre)⋅J−1(ξre))⋅Br⊥(ξre))det⁡J(ξre).(38)b_i\approx\sum_{e=1}^{n_e}\sum_{r=1}^{n_{qp}}w_r^e\left(N_i(\xi_r^e)J_z(\xi_r^e)+\nu\left(\nabla_{\xi}N_i(\xi_r^e)\cdot\mathbf{J}^{-1}(\xi_r^e)\right)\cdot\mathbf{B}^{\perp}_r(\xi_r^e)\right)\det\mathbf{J}(\xi_r^e). \tag*{(38)}

where wrew_r^e are the weights and ξre\xi_r^e are the points of a standard, nqpn_{qp}-point Gauss-Legendre quadrature rule on the parameter element Ω^e\widehat{\Omega}^e. Note that in practice, the element-wise contributions to the coefficients KijK_{ij} and bib_i are only evaluated on elements within the support of NiN_i and NjN_j, as otherwise they are zero anyway.

Importantly, the Dirichlet boundary conditions contained in the definition of the trial and test function spaces, see Eq. (9), still must be enforced on the linear system Ku=b\mathbf{K}\mathbf{u}=\mathbf{b}. In conformal IGA and FEM, this is typically done by elimination, resulting in a smaller, positive definite system to be solved. Similarly, also anti-periodic Dirichlet boundary conditions can be applied. If a weak enforcement of Dirichlet boundary conditions or of multi-domain coupling conditions using Nitsche’s method is necessary, see § 4.3 and § 4.4, the additional terms appearing in the formulations of the bilinear forms and linear forms must be discretized accordingly, which is not discussed in detail here.

Immersed boundary conformal methods

In conformal isogeometric finite element discretizations, the physical domain Ω\Omega is represented by a B-spline surface through the mapping defined in Eq. (29). However, for such a mapping F:Ω^→Ω\mathcal{F}:\widehat{\Omega}\to\Omega to exist, Ω\Omega must be topologically equivalent to a square. For most engineering applications with complex geometries and multi-material domains, this requirement is not satisfied. Within the framework of isogeometric analysis, decomposing the domain into a multi-patch geometry offers an alternative approach [14]. However, as noted previously, this process is generally cumbersome and requires significant manual effort, especially when conformity of the patches is to be maintained.

In this work, we adopt a hierarchy of three immersed isogeometric concepts to overcome this limitation:

  1. Fully immersed, see § 6.1 and [52]: Spline-based boundary representations (B-rep) of the material subdomains of the physical domain are constructed and subsequently immersed into a single background patch, the extended domain, over which the problem solution is discretized. A boundary-conformal quadrature strategy is used to accurately resolve the integration over the different regions that are trimmed from the background patch through the boundary curves of the subdomains.

  1. Union with non-conformal patches (“Union NC”), see § 6.2 and [61, 62]: To minimize the preprocessing effort, as far as desired and possible, each topologically square material subdomain of the physical domain is represented by an individual patch, which all overlay the extended domain. Thus, the overall domain consists of the union of the non-conformal patches with the trimmed regions on background patch. Nitsche’s method is used to couple non-conformal patches of the subdomains and the boundary-conformal quadrature strategy is used on the trimmed background patch.

  1. Union with conformal layers (“Union CL”), see § 6.3 and [50, 49]: In addition to the union of non-conformal patches, selected interfaces between individual patches of the subdomains and the background patch are wrapped with a conformal, overlaying patch that smoothens the transition between adjacent regions and is further trimmed from the background patch.

In the following, these three approaches are presented in more detail. Generally, we assume that the overall problem domain Ω\Omega is composed of several subdomains Ωi\Omega_i such that Ω=⋃iΩi\Omega= \bigcup_i \Omega_i and Ωi∩Ωj=∅\Omega_i \cap\Omega_j = \varnothing. For the parameterization, the background patch Ω□\Omega^{\square} is defined such that it contains the whole domain, Ω⊆Ω□\Omega\subseteq\Omega^{\square}. Each subdomain Ωi\Omega_i is parameterized either by immersion into the background patch Ω□\Omega^{\square} and trimming of the parameter domain Ω^□\widehat{\Omega}^{\square}, where the “trimmed region” corresponding to the subdomain Ωi\Omega_i is then denoted as Ω^i□\widehat{\Omega}^{\square}_i, or by an individual patch with parameter domain Ω^ip\widehat{\Omega}^{p}_i. The subdomain itself is then either obtained by the mapping of the background patch, Ωi=F□(Ω^i□)\Omega_i = \mathcal{F}^{\square}(\widehat{\Omega}^{\square}_i) or by the mapping of its individual patch, Ωi=Fip(Ω^ip)\Omega_i = \mathcal{F}^{p}_i(\widehat{\Omega}^{p}_i), respectively. Without loss of generality, in the following we assume a physical domain consisting of two disjunct material subdomains, Ω=Ω1∪Ω2\Omega= \Omega_1 \cup\Omega_2.

Fully immersed

First, we adopt the immersed IGA concept with boundary conformal quadrature, see [52]. We immerse the physical domain Ω\Omega into a background domain F□:Ω^□→Ω□\mathcal{F}^{\square} : \widehat{\Omega}^{\square} \to\Omega^{\square} such that Ω⊆Ω□\Omega\subseteq\Omega^{\square}. Each subdomain Ω1\Omega_1 and Ω2\Omega_2 is then represented in the parameter domain of the background patch by a region Ω^1□\widehat{\Omega}_1^{\square} and Ω^2□\widehat{\Omega}_2^{\square}. Each region is identified in the parametric domain by the closed loop representing its B-rep, ∂Ω^i=⋃kγ^ik\partial\widehat{\Omega}_i = \bigcup_k \widehat{\gamma}_i^k, see [63], as illustrated in Figure 2. Here, each curve γ^ik=(F□)−1(γik)\widehat{\gamma}_i^k = (\mathcal{F}^{\square})^{-1}(\gamma_i^k) represents the pull-backward of the boundary curves of the physical subdomain ∂Ωi\partial\Omega_i. While this transformation may appear cumbersome, in practice it can be trivially realized by choosing F\mathcal{F} as the identity, i.e., Ω^□\widehat{\Omega}^{\square} is identical to Ω□\Omega^{\square}, or as an affine transformation or scaling, which simply translates to an affine transformation or scaling of the control points of γik\gamma_i^k. For each region constructed through trimming the background patch Ω^□\widehat{\Omega}^{\square} with the trimming loop ∂Ω^i\partial\widehat{\Omega}_i, we accordingly restrict our mapping into that active region, hence the associated spline space over that region is defined as:

Illustration of the fully immersed concept

Figure 2. Illustration of the fully immersed concept. The physical domain Ω=Ω1∪Ω2\Omega= \Omega_1 \cup\Omega_2 is immersed into an extended domain, the background patch Ω^□\hat{\Omega}^{\square}, here consisting of 4×44 \times4 elements in its parameter domain Ω^□\hat{\Omega}^{\square}. B-reps of the subdomains Ωi\Omega_i consisting of loops of boundary curves ∂Ωi=⋃kγik\partial\Omega_i = \bigcup_k \gamma_i^k are used to determine the immersed trimmed regions Ω^i□\hat{\Omega}^{\square}_i. To determine the boundary-conformal quadrature points, the trimmed elements of each region Ω^i□\hat{\Omega}^{\square}_i are re-parameterized

N∣Ωi□=span⁡{Nj∘(F□)−1(x):j∈I, supp⁡(Nj)∩(F□)−1(Ωi□)≠∅},(39)\mathcal{N}\big|_{\Omega_i^{\square}} = \operatorname{span}\left\{N_j \circ(\mathcal{F}^{\square})^{-1}(\mathbf{x}) : j \in\mathcal{I},\ \operatorname{supp}(N_j) \cap(\mathcal{F}^{\square})^{-1}(\Omega_i^{\square}) \ne\varnothing\right\}, \tag*{(39)}

where supp⁡(Nj)⊆Ω^□\operatorname{supp}(N_j) \subseteq\widehat{\Omega}^{\square} is the support portion of the shape function Nj(ξ)N_j(\boldsymbol{\xi}). This basis function can have high continuity across material interfaces, which can cause cross-talk between different subdomains, as discontinuities cannot be correctly represented. This will be further discussed and evaluated numerically in § 7.

To achieve accurate numerical integration over the physical domain Ω\Omega, we perform the integration over each region Ω^i□\widehat{\Omega}_i^{\square} through the following procedure. Based on the representation of ∂Ωi\partial\Omega_i, the intersections between the B-rep and the vertical and horizontal knot lines of the parametric domain, i.e., the elements, are computed, see Figure 2 left and right. This operation, referred to as slicing, must be performed with sufficiently high precision, since subsequent operations, including the accuracy of the ensuing PDE analysis, are governed by this precision2. As a result of the slicing procedure, the elements of the background patch, Ω^□=⋃eΩ^e\hat{\Omega}^{\square} = \bigcup_e \hat{\Omega}^e, are categorized into three distinct groups with respect to each region, as illustrated by different colors in Figure 2, e.g. with respect to the region Ω^1□\hat{\Omega}^{\square}_1: active where ∣Ω^e∩Ω^1□∣=∣Ω^e∣|\hat{\Omega}^e \cap\hat{\Omega}^{\square}_1| = |\hat{\Omega}^e|, trimmed are those with 0<∣Ω^e∩Ω^1□∣<∣Ω^e∣0 < |\hat{\Omega}^e \cap\hat{\Omega}^{\square}_1| < |\hat{\Omega}^e| with an active portion Ω^e∩Ω^1□\hat{\Omega}^e \cap\hat{\Omega}^{\square}_1, and non-active elements with ∣Ω^e∩Ω^1□∣=0|\hat{\Omega}^e \cap\hat{\Omega}^{\square}_1| = 0, where ∣∙∣|\bullet| represents the area. The same procedure is subsequently applied to the subdomain Ω2\Omega_2, see Figure 2. Its boundary ∂Ω2\partial\Omega_2 is represented by a spline-based B-rep, and the slicing operation is performed accordingly. This results in a corresponding classification of elements into active, trimmed, and non-active categories with respect to Ω^2□\hat{\Omega}^{\square}_2.

On active non-trimmed elements, the standard Gauss-Legendre quadrature rule is used to integrate the discretized weak form, c.f. Eqs. (37) and (38). For the active portions of the trimmed elements, that are represented again through B-reps, we adopt the boundary conformal quadrature method [49]. The technique depends on a re-parameterization of the trimmed elements and consists of three stages:

  1. Approximation. For every trimmed element, the B-rep segment of the boundary of the physical domain is approximated by a Bézier curve with a polynomial degree pp of the same order as the spline discretization in Eq. (29). This allows to keep consistency under control and ensure optimal convergence for elliptic problems, as discussed in [63].

  1. Decomposition. Every trimmed element is partitioned into active and non-active regions, distinguished by the varying colors in Figure 2. When an integration cell does not conform to a standard topological triangle or quadrilateral, such as the two pentagonal active regions along the diagonal in Figure 2 (right), it is further subdivided along the trimming curve into triangular or quadrilateral sub-cells. Comprehensive details of this procedure are available in [49].

  1. Re-parametrization. In the final step, the integration cells are re-parameterized as B-spline surfaces of degree pp and quadrature points are then distributed within these cells according to a standard Gauss-Legendre rule with nqp=(p+1)2n_{qp} = (p + 1)^2 points, see Figure 2.

In this way, the boundaries ∂Ω1,∂Ω2\partial\Omega_1,\partial\Omega_2 and the integration over Ω1,Ω2\Omega_1,\Omega_2 are resolved numerically exact up to the approximation errors that stems from the re-parameterization of the trimmed cell boundaries and the quadrature order employed.

It is important to note that the decomposition and re-parameterization of trimmed elements are performed solely for the purposes of quadrature and do not alter the discretization or the number of degrees of freedom.

When employing the fully immersed technique, imposing Dirichlet boundary conditions u=gu = g on ΓD⊂∂Ω\Gamma_D \subset\partial\Omega, can be tricky if ∂Ω\partial\Omega does not coincide with ∂Ω□\partial\Omega^{\square}. In such case, Dirichlet boundary conditions cannot be imposed strongly. In our work, we impose Dirichlet boundary conditions via Nitsche’s method [33], see § 4.3.

Union with non-conformal patches

To improve the accuracy of the discretization at interfaces between subdomains, but nonetheless avoid the effort of generating a conformal multi-patch representation with only topologically square patches, we adopt the concept of overlapping, non-conformal patches, see also [61, 49, 66, 67, 68].

In Figure 3, we illustrate this method again for the physical domain consisting of two material subdomains, Ω=Ω1∪Ω2\Omega= \Omega_1 \cup\Omega_2. As in the fully immersed method, the (not topologically square) subdomain Ω1\Omega_1 is immersed into the background patch Ω□\Omega^{\square} as a trimmed region Ω^1□\widehat{\Omega}^{\square}_1. The trimming operation and the mapping back to the physical domain as Ω1=F□(Ω^1□)\Omega_1 = \mathcal{F}^{\square}(\widehat{\Omega}^{\square}_1) are conducted in accordance with § 6.1. However, now the topologically square subdomain Ω2\Omega_2 is parameterized by its own individual patch with the spline mapping F2p:Ω^2p→Ω2\mathcal{F}^{p}_{2} : \widehat{\Omega}^{p}_{2} \to\Omega_2. Thus, the active part of the background patch Ω□\Omega^{\square} corresponding to Ω1\Omega_1 is combined with the overlaying patch of Ω2\Omega_2 to construct the overall computational domain as Ω=Ω1∪Ω2\Omega= \Omega_1 \cup\Omega_2, see Figure 3.

Illustration of the union with non-conformal patches concept

Figure 3. Illustration of the union with non-conformal patches concept. As in the fully immersed method, the subdomain Ω1\Omega_1 is immersed into the background patch Ω□\Omega^{\square}. However, here Ω2\Omega_2 is parameterized by an individual patch with the mapping F2p\mathcal{F}^{p}_{2}. Thus, continuity at the interface Γ12=∂Ω1∩∂Ω2\Gamma_{12} = \partial\Omega_1 \cap\partial\Omega_2 is weakly enforced using Nitsche’s method

The fundamental concept of the union with non-conformal patches consists of weakly coupling the active parts of the background patch and the overlapping patches across their interfaces. Here, this means the continuity conditions of Eq. (17) have to be enforced on the interface Γ12=∂Ω1∩∂Ω2\Gamma_{12} = \partial\Omega_1 \cap\partial\Omega_2. As described in § 4.4, this is realized by means of Nitsche’s method [33]. Importantly, along the interface Γ12\Gamma_{12}, the discretization of the overlaying patch Ω2\Omega_2 is conformal with the interface, whereas Ω1\Omega_1 is a trimmed region of the background patch Ω□\Omega^{\square}. Thus, the weighting introduced in Nitsche’s method is always put on the untrimmed, overlaying patches as master domains, i.e., here Ω2\Omega_2, see Eq. (26). A comprehensive discussion on the construction of interface quadrature schemes used for the evaluation of interface integrals can be found in [49, 61].

Note further that while the interface conditions between the background patch and overlaying patches must be enforced weakly, the union with non-conformal patches provides the opportunity to strongly enforce Dirichlet boundary conditions on the outer boundaries of the overlaying patches, e.g., here on ΓD∩∂Ω2\Gamma_D \cap\partial\Omega_2.

Union with conformal layers

In the union with non-conformal patches, spurious effects may still occur at interfaces with strong material discontinuities, since the discretization of immersed subdomains are no conformal with the interface. To further improve the accuracy in such cases, we adopt the union with conformal layers approach, which is called “immersed boundary conformal method” in [49, 66, 50].

As in the union with non-conformal patches method, the subdomain Ω2\Omega_2 is parameterized by an individual patch with parameter domain Ω^2p\hat{\Omega}^{p}_{2} and mapping F2p\mathcal{F}^{p}_{2}. However, as illustrated in Figure 4, the subdomain Ω1\Omega_1 is now partitioned into two parts: Ω1=Ωˉ1∪ΩL\Omega_1 = \bar{\Omega}_1 \cup\Omega_L. Along the interface with Ω2\Omega_2, the new subdomain ΩL\Omega_L forms a geometrically conformal layer, which is generated by extruding Γ12\Gamma_{12} into Ω1\Omega_1, yielding a new interface curve Γ1L\Gamma_{1L}. This subdomain is also parameterized by an individual patch with the mapping FLp:Ω^Lp→ΩL\mathcal{F}^{p}_{L} : \hat{\Omega}^{p}_{L} \to\Omega_L and maintains the same material parameters as Ω1\Omega_1. The cropped subdomain Ωˉ1=Ω1\ΩL\bar{\Omega}_1 = \Omega_1\backslash\Omega_L remains immersed into the background domain as region Ωˉ^1□\hat{\bar{\Omega}}^{\square}_1. It is defined by a modified set of boundary curves ∂Ωˉ1=⋃kγ^1k\partial\bar{\Omega}_1 = \bigcup_k\hat{\gamma}^{k}_1, which is determined from the original boundary ∂Ω1\partial\Omega_1 through the extrusion operation that establishes ΩL\Omega_L.

Illustration of the union with conformal layers concept

Figure 4. Illustration of the union with conformal layers concept. As in the union with non-conformal patches method, the subdomain Ω2\Omega_2 is parameterized by an individual patch with the mapping F2p\mathcal{F}^{p}_{2}. Beyond, here the subdomain Ω1\Omega_1 is partitioned into two parts: Along the interface with Ω2\Omega_2, the new subdomain ΩL\Omega_L forms a geometrically conformal layer, which is offset from the boundary Γ12\Gamma_{12} and also parameterized by an individual patch with the mapping FLp\mathcal{F}^{p}_{L}. The cropped subdomain Ωˉ1=Ω1\ΩL\bar{\Omega}_1 = \Omega_1\backslash\Omega_L remains immersed in the background domain as region Ωˉ^1□\hat{\bar{\Omega}}^{\square}_1, which is defined by a modified set of boundary curves ∂Ωˉ^1=⋃kγ^1k\partial\hat{\bar{\Omega}}_1 = \bigcup_k\hat{\gamma}^{k}_1. Continuity at the two interfaces Γ1L=∂Ωˉ1∩∂ΩL\Gamma_{1L} = \partial\bar{\Omega}_1 \cap\partial\Omega_L and ΓL2=∂ΩL∩∂Ω2\Gamma_{L2} = \partial\Omega_L \cap\partial\Omega_2 is again weakly enforced using Nitsche’s method

Here, continuity at the two interfaces Γ1L=∂Ωˉ1∩∂ΩL\Gamma_{1L} = \partial\bar{\Omega}_1 \cap\partial\Omega_L and ΓL2=∂ΩL∩∂Ω2\Gamma_{L2} = \partial\Omega_L \cap\partial\Omega_2 is again weakly enforced using Nitsche’s method, as described in § 4.4. For Γ1L\Gamma_{1L}, ΩL\Omega_L must be the master domain, as it is discretized conformal with the interface, whereas Ωˉ1\bar{\Omega}_1 is trimmed. For ΓL2\Gamma_{L2}, either patch could be the master, as both ΩL\Omega_L and Ω2\Omega_2 are geometrically conformal. We emphasize, however, that conformity is understood purely in a geometric sense: the boundary layer ΩL\Omega_L inherits the geometric shape of the interface Γ12\Gamma_{12}, but no constraint is imposed on the discrete function spaces. In particular, the discretization of ΩL\Omega_L is not required to be conforming with that of Ω2\Omega_2, i.e., the degrees and number of elements along the interface direction may vary in each patch. Thus, a weak enforcement of the interface conditions is also applied on ΓL2\Gamma_{L2}.

Preconditioning and stability

The application of immersed methods in FEM/IGA is often associated with stability issues and ill-conditioning [69]. Although these issues are closely related, their treatment generally requires different techniques.

Both issues are directly related to the small cut element problem, where elements are trimmed, leaving only a tiny active part. In the context of stability, this becomes evident when Dirichlet boundary conditions need to be imposed on a boundary that trims the background domain, resulting in tiny active parts. In such configurations, the weak imposition of Dirichlet boundary conditions via, e.g., Nitsche’s method leads to an ill-posed numerical formulation, requiring a penalty parameter that tends to infinity to maintain coercivity of the formulation. In the context of immersed and trimmed IGA, several unstable configurations were investigated, including cases where trimming produces high aspect ratio rectangular trimmed elements [70]. In such cases, instabilities were observed to be associated with the flux term. In the framework of unfitted FEM, similar stability issues have been reported, and the reader is referred to [71]. Additional discussions on stability aspects in immersed and trimmed spline methods can be found in [72, 73, 74, 75, 76]. However, in the present work, no additional stabilization technique is required in the union methods, as the weighted flux term in Nitsche’s method is evaluated exclusively from the untrimmed domain, see Eq. (26) in § 4.4, as discussed in [61].

Furthermore, ill-conditioning in immersed methods arises due to the small cut element problem, where some basis functions lose a significant portion of their support, leading to near-linear dependencies in the discrete system. Detailed discussions on this phenomenon can be found in [77, 78, 79, 80]. In the present work, diagonal scaling in the form of Jacobi preconditioning is applied to improve the conditioning of the system matrix. Furthermore, a sparse direct solver is used in MATLAB, and no conditioning issues are observed in the numerical investigation.

Numerical results

In the following, we apply immersed IGA with the three aforementioned concepts to two-dimensional magnetostatic problems to demonstrate their efficiency and accuracy. The current work is implemented within the framework of the open-source MATLAB code geoPDEs [81, 82] and an in-house library based on irit [64] and Open CASCADE [65] for obtaining the boundary conformal quadrature points.

Throughout this manuscript, the “kk-version” of IGA is adopted as the default refinement strategy, i.e., kk-refinement is used by increasing the smoothness at internal knots along with the degree [11]. Any deviation from this approach, such as the use of the FEM pp-refinement, will be explicitly stated where applicable.

In the following, we evaluate the accuracy of the obtained numerical solutions uhu^h against reference solutions urefu^{ref}, which could be either analytical or highly-refined multipatch IGA solutions, using the L2L^2-norm and the H1H^1-seminorm of the error, defined as follows:

∥uh−uref∥L22=∫Ω∥uh−uref∥22 dΩ,∥uh−uref∥H12=∫Ω∥∇uh−∇uref∥22 dΩ.(40)\begin{aligned} \lVert u^h-u^{ref}\rVert_{L^2}^2 &= \int_{\Omega}\lVert u^h-u^{ref}\rVert_2^2\,\mathrm{d}\Omega,\\ \lVert u^h-u^{ref}\rVert_{H^1}^2 &= \int_{\Omega}\lVert\nabla u^h-\nabla u^{ref}\rVert_2^2\,\mathrm{d}\Omega. \tag*{(40)} \end{aligned}

Coaxial cable

First, we consider a coaxial cable as benchmark problem, see Figure 5a and [47, 55, 83] for similar studies. The geometry consists of three concentric cylindrical regions: an inner conducting core (gold), a surrounding insulating layer (green), and an outer conductor (orchid). Due to the double symmetry of the problem, we only solve quarter of the cable geometry. The core has radius r1r_1, the insulator occupies the annular region r1<r<r2r_1 < r < r_2, and the outer conductor extends for r2<r<r3r_2 < r < r_3. Here, the radii are specified as r1=13r_1=\frac{1}{3}, r2=23r_2=\frac{2}{3}, r3=1r_3=1. An axial current flows in the core in the positive zz-direction (out of plane) with magnitude I1=1000I_1=1000. The outer conductor carries an equal current in the opposite direction, I3=−I1I_3=-I_1. We assume a unified permeability for the three materials, equal to the one of air μ=μ0=4π×10−7 H/m\mu=\mu_0=4\pi\times10^{-7}\,\mathrm{H/m}. It is important to note that the remanence term in Eq. (16) is omitted, since all materials are assumed to be non-magnetic, i.e., Br=0\mathbf{B}_r=0.

Illustration of the coaxial cable problem setup

Figure 5. Illustration of the coaxial cable problem setup

A Dirichlet boundary condition is imposed on the outer radius of the cable r=r3r=r_3 with Az=0A_z=0. No additional boundary conditions are needed at the left and bottom material boundaries. There, the magnetic field is perpendicular to the surface, which implies the natural Neumann condition ∇Az⋅n=0\nabla A_z\cdot\mathbf{n}=0. This condition is inherently satisfied in the weak formulation, see Eqs. (10) and (13).

The analytical solution for the magnetic flux density is formulated in terms of the radial coordinate r=x2+y2r = \sqrt{x^2+y^2}, see [47, 55], as follows:

Bθ(r)={μ0I2πr12r,0≤r<r1,μ0I2πr,r1≤r<r2,μ0I(r32−r2)2πr(r32−r22),r2≤r<r3,0,r≥r3.(41)B_\theta(r)= \begin{cases} \frac{\mu_0 I}{2\pi r_1^2}r, & 0 \le r < r_1, \\ \frac{\mu_0 I}{2\pi r}, & r_1 \le r < r_2, \\ \frac{\mu_0 I(r_3^2-r^2)}{2\pi r(r_3^2-r_2^2)}, & r_2 \le r < r_3, \\ 0, & r \ge r_3. \end{cases} \tag*{(41)}

The scalar potential Az:=urefA_z := u^{ref} can be obtained by integrating Eq. (41) with integration constants determined by continuity conditions.

Az(r)=−∫0rBθ(ρ) dρ.(42)A_z(r)=-\int_0^r B_\theta(\rho)\,\mathrm{d}\rho. \tag*{(42)}

First, we solve the coaxial problem using the fully immersed approach, see § 6.1. The physical domain Ω\Omega is embedded into an extended square domain Ω□=[0,r3]2\Omega^\square=[0,r_3]^2. Similar to the example shown in Figure 2, the present physical domain consists of multiple subregions. Accordingly, a separate trimming loop is defined for each subdomain Ωi\Omega_i. Each trimming loop is applied to the extended domain Ω□\Omega^\square, resulting in the identification of the corresponding active and trimmed elements associated with each subregion, in addition to the corresponding integration points for trimmed elements and, at each quadrature point on the boundary, the corresponding outward normal vector. Finally, we impose weakly the Dirichlet boundary condition Az=0A_z=0 on the outer radius r=r3r=r_3 via Nitsche’s method as this domain boundary does not coincide with the boundary ∂Ω□\partial\Omega^\square. Furthermore, the integration domain is restricted exclusively to the physical domain Ω\Omega, while all degrees of freedom associated solely with the fictitious domain Ω□\Ω\Omega^\square\backslash\Omega (the gray region shown in Figure 5b) are excluded from the system. By only integrating over the active part of the trimmed elements, there will be some basis functions with very small support, which may result in an ill-conditioned system matrix. To alleviate this issue, a diagonal preconditioner is employed.

Second, we solve the problem using union with non-conformal patches approach, following § 6.2. In Figure 5c, the physical domain Ω\Omega is defined as the union of the overlapping patches, Ω=Ω1□∪Ω2p∪Ω3p\Omega=\Omega_1^\square\cup\Omega_2^p\cup\Omega_3^p, where Ω1□\Omega_1^\square denotes the active portion of the trimmed patch Ω1\Omega_1. Here, we impose the Dirichlet boundary condition Az=0A_z=0 strongly via elimination.

For both methods, the fully immersed and union with non-conformal patches, we set the stabilization parameter in Nitsche’s method to be βνairp/h\beta\nu_{\mathrm{air}}p/h, where β∈{100,1000}\beta\in\{100,1000\} and hh denotes the smallest element size in the domain.

The hh-refinement analysis at different polynomial degrees in Figure 6a and Figure 6b shows optimal convergence rates for both the L2L^2-norm error and the Hs1H_s^1-seminorm error for the Union NC method. However, for the fully immersed case, the pp-convergence behavior is pre-asymptotic and hence the convergence rates for p>1p > 1 are not optimal. This can be attributed to the different inter-patch continuity properties. In the union formulation, only C0C^0 continuity is enforced weakly across the patch interfaces. By contrast, in the fully immersed approach, both the solution and its gradients (i.e., the flux density) are represented by basis functions exhibiting Cp−1C^{p-1} continuity across the regions’ interfaces, as we are solving on a single patch. This explains the similarity in accuracy for the case p=1p = 1, as the gradients are represented by non-smooth basis functions. This can be seen clearly in Figure 6c and Figure 6d, where the errors are nearly identical in terms of the Hs1H_s^1-seminorm and remain very close in the L2L^2-norm for the case p=1p = 1. Overall, the union approach provides higher accuracy for higher polynomial orders while maintaining a comparable number of degrees of freedom.

Comparison of convergence errors for immersed and union non-conformal patch methods

Figure 6. Comparison of convergence behaviors of the fully immersed and union with non-conformal patches methods over the element size hh and the average number of degrees of freedom DOFs\sqrt{\mathrm{DOFs}} for the coaxial cable problem

Additionally, Figure 7a and Figure 7b shows the contours of the scalar potential AzA_z and the magnitude of the flux density BθB_\theta for a discretization with degree p=2p = 2 and 32×3232 \times32 elements for each patch for the union method. To further assess the performance, we compare the exact solution along the line y=xy = x with the corresponding numerical results obtained using both methods. Figure 7 demonstrates excellent agreement between the analytical and numerical solutions already for this coarse, low-degree discretization for the fully immersed and union of non-conformal patches methods. However, as the insets show, the kinks of BθB_\theta at the interfaces are less well resolved by the fully immersed approach, which may explain the lack of higher-order convergence.

Contours of scalar potential $A_z$ and magnitude of flux density $B_\theta$, with comparisons along $x=y$

Figure 7. Comparison of the scalar potential AzA_z and the magnitude of the flux density BθB_\theta between the analytical solution and the numerical results obtained using the fully immersed and union with non-conformal patches methods with p=2p=2 and 32×3232 \times32 elements for each patch

Horseshoe magnet

To further assess the performance of the three proposed methods on a more complex problem, we consider a straight horseshoe magnet. The problem consists of two soft iron rods, one of which is connected to two permanent magnets (PMs) having opposite remanence directions, as illustrated in Figure 8. The computational domain has a width of W=0.15W = 0.15 mm and a height of H=0.13H = 0.13 mm. Each soft iron rod has a width of Wiron=0.02W_{\mathrm{iron}} = 0.02 mm and a height of Hiron=0.09H_{\mathrm{iron}} = 0.09 mm. The permanent magnets each have dimensions WPM=0.05W_{\mathrm{PM}} = 0.05 mm and HPM=0.02H_{\mathrm{PM}} = 0.02 mm. The magnitude of the remanence of each magnet is prescribed as Br=1.4B_r = 1.4. The relative permeability of the permanent magnets is μrPM=1.05\mu_r^{\mathrm{PM}} = 1.05, such that their permeability is given by μPM=μrPMμ0\mu_{\mathrm{PM}} = \mu_r^{\mathrm{PM}}\mu_0, where μ0\mu_0 denotes the vacuum permeability. The soft iron rods are modeled with a relative permeability of μriron=4000\mu_r^{\mathrm{iron}} = 4000, while the surrounding air region is assigned the vacuum permeability, i.e., μair=μ0\mu^{\mathrm{air}} = \mu_0. A Dirichlet boundary condition is defined on the outer boundary of the computational domain with Az=0A_z = 0.

Image not extracted

Figure 8. Schematic drawing of the horseshoe magnet problem. The red rectangular regions denote permanent magnets with different magnetization directions BrB_{\mathrm{r}}. The gray rectangular regions correspond to soft iron rods. The surrounding blue domain represents the air region

Here, we compare all three immersed IGA methods, see Figure 9 for illustrations of the meshes. To have a reference solution urefu^{\mathrm{ref}}, we solve the horseshoe magnet problem with conformal multi-patch IGA using 30 conformal patches with each patch being discretized with 20×2020 \times20 elements and polynomial degree p=5p = 5, with a total of 17,545 DOFs, see Figure 9a. To compare the numerical solution of the fully immersed approach with the conformal multi-patch solution, we evaluate the error norms, see Eq. (40), as well as the magnitude of the flux density over the horizontal line y=0.1y = 0.1, which is indicated in blue in Figure 8.

Image not extracted

Figure 9. Schematic drawings for the discretizations of the horseshoe magnet problem with standard multi-patch IGA and the three proposed immersed IGA methods

First, we start with the fully immersed approach, where trimming loops are defined for each soft iron rod and permanent magnet component. We follow the same procedure as in § 6.1, where every active and trimmed element corresponding to a specific region is classified.

The system matrix and RHS vector, see Eqs. (37) and (38), are then obtained by integrating over the corresponding elements for all the different regions/components (magnets, iron rods, air).

The results shown in Figure 10 demonstrate that, for both kk- and pp-refinement strategies, with a background grid of 128×128128 \times128 knot spans, the magnetic flux density magnitude is largely consistent with the reference solution within each material subdomain and sufficiently far from the material interfaces. However, for the cases employing kk-refinement with highly smooth basis functions with Cp−1C^{p-1}-continuity across knot spans, a significant discrepancy compared to the reference solution arises in the vicinity of the material interfaces. In particular, the lowest polynomial order p=1p = 1 exhibits the smallest deviation from the reference solution relative to the higher orders p={2,5}p = \{2,5\}. This behavior can be explained by the low continuity of the basis functions at p=1p = 1, where the gradients are discontinuous across element boundaries. Since the magnetic flux density is obtained as the spatial derivative of the magnetic potential, this discontinuity enables a more accurate representation of the magnetic flux density at material interfaces. In contrast, higher polynomial orders combined with high inter-element continuity enforce smoothness of both the solution and its derivatives, i.e., also the magnetic flux density, across knot spans. As a consequence, the magnetic flux density is approximated by overly smooth basis functions across interfaces where discontinuities in material properties occur. This leads to spurious oscillations characteristic of Gibbs-type phenomena, as observed in Figure 10. These spurious oscillations can be mitigated by employing pp-refinement with C0C^{0}-continuity, where the continuity across knot spans is reduced to C0C^{0}.

Image not extracted

Figure 10. Evaluation of the magnetic flux density magnitude along the line y=0.1y = 0.1 for different polynomial orders p=1,2,5p = 1, 2, 5 for both kk- and pp-refinement strategies with a background grid of 128×128128 \times128 knot spans for the fully immersed method

The lower inter-element continuity allows for a more flexible representation of gradient variations at material interfaces, thereby reducing non-physical oscillatory behavior. Nevertheless, if a material interface is located within the interior of an element rather than coinciding with element boundaries, spurious oscillations may still arise, as observed in Figure 10 in the magnifiers at x=0.02x = 0.02 and x=0.13x = 0.13. In such cases, the approximation space remains smooth within the element, preventing an accurate representation of the discontinuity in the material coefficients and consequently leading to localized oscillatory artifacts. Thus, it is evident that the fully immersed strategy is not suitable for accurately handling multiple materials within the same patch.

To mitigate the fictitious oscillations over the material interfaces, we employ the union with non-conformal patches approach and the union with conformal layers approach. In the former approach, we couple four non-conformal patches alongside the trimmed background patch, following the procedures in § 6.2, as illustrated in Figure 9c. For the latter method, the non-conformal patches are further enclosed with two conformal layers, as introduced in § 6.3 and illustrated in Figure 9d. The results in Figure 11 show a comparison of the evaluation of the magnetic flux density over the line y=0.1y = 0.1 for both methods with the reference solution. The union NC numerical solution is evaluated with a polynomial degree of p=5p = 5 and each patch is discretized with uniform element size h=0.01h = 0.01, resulting in a total number of 13,288 DOFs. For the union CL solution, we use the same elements size and polynomial degree, resulting in a slightly higher total number of 16,638 DOFs. Here, it can clearly be observed that both methods align perfectly with the multi-patch IGA evaluation. It is evident that both methods mitigate the fictitious oscillations over the interfaces at a high polynomial degree, even with a smaller number of DOFs compared with the reference solution, which has 17,545 degrees of freedom.

Image not extracted

Figure 11. Evaluation of the magnetic flux density magnitude along the line y=0.1y = 0.1 for polynomial order p=5p = 5 for the union with non-conformal patches approach and the union with conformal layers method with two conformal layers, both with element size h=0.01h = 0.01

To further evaluate the performance of both approaches, we perform an hh-ref analysis study for both methods for different polynomial orders p={2,4,6}p = \{2, 4, 6\}. The stabilization parameter is set to βνmax⁡p/h\beta\nu_{\max}p/h, where β=104\beta= 10^{4}, νmax⁡\nu_{\max} denotes the maximum reluctivity (corresponding to air in the present configuration), pp is the polynomial order, and hh corresponds to the element size, as every patch is discretized with uniform element size hh. The L2L^{2}-norm and H1H^{1}-seminorm error measures as defined in Eq. (40) are considered in comparison to the conformal multi-patch IGA reference solution, as the horseshoe magnet problem has no exact solution. For the calculations of the error norms, a (p+10)×(p+10)(p + 10) \times(p + 10) Gauss-Legendre quadrature rule is used within each element. The results shown in Figure 12 indicate that, for the union NC method, mesh refinement through knot insertion improves accuracy in both the L2L^{2}-norm and the H1H^{1}-seminorm errors. In contrast, increasing the polynomial order does not lead to a noticeable improvement in accuracy for this approach. For the union CL method, however, the error decreases with increasing polynomial order, indicating a more effective exploitation of higher-order basis functions. Nevertheless, neither method achieves the optimal theoretical convergence rate. This behavior can be attributed to geometric discontinuities at the rectangles’ corners, which reduce the regularity of the solution. Consequently, the convergence rate is limited by the reduced solution regularity rather than by the order of the basis functions.

Image not extracted

Figure 12. Comparison of convergence behaviors of the union with non-conformal patches and the union with conformal layers methods for the horseshoe magnet problem

Finally, Figure 13 shows the contours of the scalar potential field AzA_z and the magnitude of the magnetic flux density ∥B∥\lVert\mathbf{B} \rVert for the conformal multi-patch IGA reference solution, as well as the union methods. The union CL demonstrate an improved capability to capture singular field behavior compared to the union NC method. In particular, the singularities occurring at the corners of the left boundary of the right iron rod are clearly resolved in Figure 13b and Figure 13f. In contrast, these localized singular features are not adequately captured by the union NC method, as observed in Figure 13d.

Image not extracted

(a) AzA_z for reference multi-patch IGA

Image not extracted

(b) ∥B∥\|\mathbf{B}\| for reference multi-patch IGA

Image not extracted

(c) AzA_z for union with non-conformal patches (h=0.01h = 0.01)

Image not extracted

(d) ∥B∥\|\mathbf{B}\| for union with non-conformal patches (h=0.01h = 0.01)

Image not extracted

(e) AzA_z for union with conformal layers (h=0.01h = 0.01)

Image not extracted

(f) ∥B∥\|\mathbf{B}\| for union with conformal layers (h=0.01h = 0.01)

Figure 13. Contours of scalar potential (left) and magnetic flux density (right) for horseshoe magnet problem

Permanent magnet assembly for a magnetocaloric cooling device

As a final benchmark, we apply the union methods to an industry-related problem: the simulation of a permanent magnet assembly of a magnetocaloric cooling device. Magnetocaloric systems can adopt different design configurations; in this study, we consider a co-rotary configuration consisting of an inner rotor and an outer rotor separated by an air gap (see Figure 14). The air gap typically hosts the active magnetic regenerator units, where heat exchange occurs between the working fluid (e.g., water) and the magnetocaloric material [84].

Image not extracted

Figure 14. Schematic representation of the permanent magnet assembly. The permanent magnets are shown in red, the iron components in gray, and the air regions in light blue. Each magnet a magnetization direction with θ=45∘\theta= 45^\circ measured from the xx-axis

In the present work, however, the air gap is assumed to contain only air [85]. Consequently, assessing the cooling performance of such a device requires an accurate evaluation of the magnetic field along a fixed radius within the air gap [86, 85]. In this study, the analysis is limited to the accurate evaluation of the magnetic field distribution, whereas a subsequent thermodynamic performance analysis lies beyond the scope.

In the following, we briefly describe the main geometric dimensions of the design. The permanent magnet assembly shown in Figure 14 consists of an inner rotor with an inner radius of r1=30r_1 = 30 mm and an outer radius of r2=130r_2 = 130 mm, as well as an outer rotor with an inner radius of r3=155r_3 = 155 mm and an outer radius of r4=300r_4 = 300 mm. Both permanent magnets have a dimension of 6060 mm ×\times 6060 mm. The flux collector in the outer rotor has a height of HORFC=5H_{\mathrm{ORFC}} = 5 mm, while it is HIRFC=15H_{\mathrm{IRFC}} = 15 mm for the inner rotor. Dirichlet boundary conditions Az=0A_z = 0 are imposed strongly by elimination on the inner radius of the inner rotor and on the outer radius of the outer rotor, which physically implies that no magnetic flux crosses these boundaries. Additionally, an anti-periodic boundary conditions are imposed on the sides Γ+\Gamma^{+} and Γ−\Gamma^{-}. The material properties of the permanent magnets, iron, and air are the same as those used in the horseshoe magnet problem, see § 7.2. For comparison, we solve the problem using conformal multi-patch IGA with 21 conformal patches, each being discretized with 30×3030 \times30 knot spans and polynomial order of p=2p = 2,

resulting in a total number of 20,492 DOFs. We evaluate the magnetic flux density over the mean radius between both rotors at ra=142.5r_a = 142.5 mm, as highlighted by the blue dashed curve in Figure 14.

For illustration purposes, examples of the spatial discretization for the considered approaches are shown in Figure 15. In Figure 15a, the geometry is parametrized using 21 conformal patches, each discretized with 4×44 \times4 knot spans. Constructing such a parametrization is a cumbersome process, as the corner points of each patch must be carefully identified and conformity between adjacent patches must be ensured. In contrast, for the union with non-conformal patches approach shown in Figure 15b, the geometry can be parametrized by unifying six non-conformal patches over a trimmed background patch, which significantly reduces the parametrization effort. For the union with conformal layers, two configurations are considered for constructing the conformal layers. In the first configuration, see Figure 15c, the conformal layer is constructed only along the outer boundaries of the flux collectors in the inner and outer rotors. In the second configuration, see Figure 15d, the conformal layers wrap around the flux collectors and part of the permanent magnets in both the inner and outer rotors. The total number of patches for both configurations is nine (eight non-conformal patches unified with a trimmed background patch).

Different geometric discretizations of the permanent magnet assembly, showing (a) Multi-patch IGA, (b) Union with non-conformal patches, (c) Union with conformal layers (version 1), and (d) Union with conformal layers (version 2)

Figure 15. (a) Multi-patch IGA; (b) Union with non-conformal patches; (c) Union with conformal layers (version 1); (d) Union with conformal layers (version 2). Illustration of the different geometric discretizations of the permanent magnet assembly

For the union with non-conformal patches approach, the number of elements used for the discretization of each patch is shown in Table 1. The total number of degrees of freedom is 23,869. For the union with conformal layers configurations, the total number of degrees of freedom is 25,621 for version 1 and 24,477 for version 2. The conformal layers in version 2 are constructed in such a way to allow the solution field to be resolved more accurately in the vicinity of the singularities. Consequently, a coarser background mesh is used in this configuration, resulting in fewer degrees of freedom compared to version 1, see Table 1. For all three methods, basis functions with a polynomial degree of p=2p = 2 are used, and the stabilization parameter for Nitsche’s method was set to β/h=5⋅1011\beta/h = 5 \cdot10^{11}.

PatchUnion NCUnion CL (v1)Union CL (v2)
Iron core20×3020 \times3020×3020 \times3020×3020 \times30
PM140×4040 \times4040×4040 \times4040×4040 \times40
Iron IRFC20×4020 \times4020×4020 \times4020×4020 \times40
Iron ORFC20×4020 \times4020×4020 \times4020×4020 \times40
PM240×4040 \times4040×4040 \times4040×4040 \times40
Iron bridge50×5050 \times5050×5050 \times5050×5050 \times50
Boundary layers20×4020 \times4010×8010 \times80
Air (Background)100×100100 \times100100×100100 \times10060×6060 \times60

Table 1. Space discretization for each approach in terms of knot spans

The evaluation of the magnitudes of the flux field ∥B∥\lVert\mathbf{B}\rVert at ra=142.5r_a = 142.5 mm shown in Figure 16 indicates that the union with non-conformal patches approach yields the lowest accuracy, with the largest discrepancies occurring in the high field region between the two flux collectors (inner and outer rotors), where the flux density is overestimated. Along the circumferential angle θ\theta, the field is generally underestimated, except near both ends of the profile, where the solution again exhibits overestimation. Overall, these trends suggest that the union NC approach requires a finer discretization of the trimmed background patch to achieve a more accurate field evaluation.

Evaluation of flux density magnitude in the air gap of the permanent magnet assembly

Figure 16. Evaluation of flux density magnitude in the air gap of the permanent magnet assembly at ra=142.5r_a = 142.5 mm

For version 1 of the union with conformal layers, the predicted field closely matches the multi-patch IGA reference solution, with only a very slight overestimation in the angular interval θ∈[35,55]\theta\in[35, 55]. This deviation can be reduced by additional refinement of the conformal boundary layers. Finally, the version 2 configuration exhibits an almost perfect agreement with the multi-patch IGA solution, despite employing fewer degrees of freedom than version 1.

To finalize the comparison, the contours of the magnetic flux density magnitude ∥B∥\lVert\mathbf{B}\rVert are shown in Figure 17. For the union NC approach, the contours generally agree with those of the reference solution. However, small but noticeable discrepancies are observed within the air gap and in the vicinity of the flux collectors. As discussed earlier, these discrepancies could be mitigated by further refinement of the background patch. In contrast, both union CL configurations show excellent agreement with the reference contours throughout the domain.

Contours of magnetic flux density of the permanent magnet assembly for different methods

Figure 17. Contours of magnetic flux density of the permanent magnet assembly for different methods

Discussion

Overall, the three examples presented pose different challenges to the three immersed IGA approaches adopted in this work (fully immersed, union with non-conformal patches, and union with conformal layers).

The coaxial cable problem introduces a discontinuous source current density across different subdomains. Here, the fully immersed approach is able to accurately predict the magnetic potential and the flux density within each subdomain. However, it cannot capture the flux density profile across subdomain interfaces exactly. This is due to the inability to represent the discontinuities in the solution gradient through smooth basis functions across the interfaces. On the contrary, the Union NC approach accurately resolves the flux density over subdomain interfaces. In addition, the approach exhibits optimal convergence rates in the hh-refinement analysis for different polynomial orders, unlike the fully immersed approach.

The horseshoe magnet problem introduces a different challenge compared to the coaxial cable problem: discontinuous material parameters across several subdomains. The fully immersed approach with kk-refinement was able to represent the scalar potential and the magnetic flux density within each material well. However, as the magnetic flux density, i.e., a gradient field, is approximated with highly continuous basis function, Gibbs-type phenomena are observed at material interfaces. Although this can be reduced when pp-refinement is employed, the Gibbs-type phenomena can still be observed within the first few knot spans around the interfaces. For the Union NC approach, these oscillations were suppressed, as each material is represented by an independent patch and thus independent basis functions. However, the method requires refinement, mainly near singularities, to accurately capture them. This is possible with the implementation of the Union CL method, where critical areas are wrapped with conformal layers to accurately capture local features and accurately resolve the magnetic flux density. This can be seen in particular at the corners of the iron rods, where singularities occur, indicated by spikes in the high flux density. Overall, the Union CL has shown to have less error in both L2L^2-norm and Hs1H_s^1-seminorm by approximately one order of magnitude.

The permanent magnet assembly problem features similar challenges as the horseshoe magnet problem, however, the geometry is more complex. Here, the industry-related target was to accurately evaluate the magnetic field within the air gap. This requires an accurate representation of the geometry and ensuring that no cross-talk occurs between different materials/subdomains. The Union NC and Union CL approaches were examined against a reference solution obtained by conventional conformal multi-patch IGA. The Union NC was found to overestimate the field at the high-field region, and slightly underestimate the field near the low-field region. However, with two different boundary conformal configurations, the Union CL was able to accurately evaluate the gap field.

Conclusions and future work

In this work, we developed an isogeometric immersed boundary-conformal framework for two-dimensional magnetostatic problems, in which three different methods are employed, namely: fully immersed, union with non-conformal patches, and union with conformal layers. In all three approaches, trimmed elements in the immersed background patch are efficiently handled by adopting a boundary-conformal quadrature method, leading to accurate integration with a low number of quadrature points. Coupling non-conformal patches is enabled by employing Nitsche’s method while ensuring stability by considering the flux from the non-trimmed patches, which alleviates the need for additional stabilization methods.

The results demonstrate that the proposed framework achieves accuracy comparable to conformal multi-patch IGA while significantly reducing the geometry preprocessing effort. We investigated the application of the aforementioned methods for three different problems, each presenting distinct challenges. The coaxial cable problem is intended to represent a discontinuous current field between several regions. For the horseshoe magnet problem, discontinuous material parameters, i.e., permeability, introduce difficulties in treating Gibbs-type phenomena. In the last example, the geometry represents a cross-section of an industrial application, where accurate evaluation of the field in the air gap plays a crucial role in predicting the cooling power of the device.

The proposed immersed IGA framework with non-conformal patches and conformal layers offers several advantages compared to conventional frameworks, e.g., multi-patch IGA:

  • By employing the union approaches, multiple subdomains can be described directly by immersing their B-rep – which can be extracted from CAD – and conformal layers can then be generated automatically without human intervention. This is not the case for standard multi-patch IGA, where conforming patches must be constructed manually.

  • The geometric preprocessing effort and the number of patches required are significantly reduced, e.g., from 30 to 5 in the horseshoe magnet problem and from 21 to 7 for the permanent magnet assembly.

  • The method circumvents the requirement for conformity between independent patches, unlike multi-patch IGA, thereby reducing the effort associated with generating a conformal multi-patch mesh. This also applies to weakly coupled multi-patch IGA based on Nitsche’s method [34]. However, such approaches are typically limited to regions that are topologically equivalent to a square, e.g., making it difficult to represent the air region in § 7.3 using a single patch.

  • The framework removes the need for Lagrange multiplier methods or harmonic mortaring methods to couple different subdomains, resulting in a system with fewer unknowns.

  • The union concept (used in the Union NC and Union CL methods) allows for strong imposition of Dirichlet boundary conditions on domain boundaries, unlike the fully immersed fictitious domain procedure, where this is only possible when the physical domain boundaries coincide with the extended domain boundaries.

  • The Union CL enables accurate capture of solution singularities, resulting in improved accuracy.

  • The Union CL demonstrates its effectiveness in mitigating spurious oscillations across material interfaces due to the construction of a conformal layer within the trimmed field, which leads to smoother transitions across interfaces.

  • Both the Union NC and Union CL provide accurate evaluation of the field within critical regions (e.g., the air gap in § 7.3).

Overall, the union with non-conformal patches method provides accurate results for the coaxial cable problem. However, for the horseshoe magnet problem and the permanent magnet assembly problem, the method requires refinement near critical regions to mitigate over- or underestimation of the field. In contrast, the union with conformal layers provides highly accurate results for all considered problems, demonstrating its accuracy, efficacy, and robustness for two-dimensional magnetostatic simulations.

Future work will investigate the effect of local refinement strategies, such as Truncated Hierarchical B-splines [87, 28, 88], particularly near critical regions, which may reduce the number of conformal layers required. Furthermore, the method will be extended to three-dimensional magnetostatic problems and applied to additional electric machine configurations [89].

Acknowledgments

Y.T.E. and O.W. acknowledge financial support by funds from LOEWE – the Hessian State Offensive for the Development of Scientific and Economic Excellence (Landes-Offensive zur Entwicklung Wissenschaftlich-ökonomischer Exzellenz, Förderlinie 3: KMU-Verbundvorhaben), HA project no. 1450/23-04, and of the Graduate School Computational Engineering at the Technical University of Darmstadt, Germany.

G.G. and P.A. acknowledge the financial support of the Swiss National Science Foundation through the project FLASH with no. 200021_214987.

The authors would like to thank Mr. Daniel Salazar for his assistance in performing the simulation of the horseshoe magnet with the fully immersed method.

CRediT authorship contribution statement

Yusuf T. Elbadry: Writing – original draft, Visualization, Validation, Software, Investigation, Conceptualization; Giuliano Guarino: Writing – review & editing, Visualization, Validation, Software, Conceptualization; Pablo Antolín: Writing – review & editing, Software, Methodology, Conceptualization; Oliver Weeger: writing—review & editing, Visualization, Methodology, Supervision, Resources, Funding acquisition, Conceptualization.

Data Availability

None.

Disclosure Statement

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Declaration of generative AI and AI-assisted technologies in the writing process

The authors disclose the use of ChatGPT, a large language model developed by OpenAI [90], as a language editing tool to enhance the clarity and readability of the manuscript. The authors are solely responsible for the content of this work.

References

  1. [1]J. Baiges, R. Codina, Variational Multiscale error estimators for solid mechanics adaptive simulations: An Orthogonal Subgrid Scale approach, Computer Methods in Applied Mechanics and Engineering 325 (2017) 37–55. doi:10.1016/j.cma.2017.07.008.DOI
  2. [2]D. Codony, O. Marco, S. Fernández-Méndez, I. Arias, An immersed boundary hierarchical B-spline method for flexoelectricity, Computer Methods in Applied Mechanics and Engineering 354 (2019) 750–782. doi:10.1016/j.cma.2019.05.036.DOI
  3. [3]J. C. Alzate Cobo, X.-L. Peng, B.-X. Xu, O. Weeger, A finite swelling 3D beam model with axial and radial diffusion, Computer Methods in Applied Mechanics and Engineering 441 (2025) 117983. doi:10.1016/j.cma.2025.117983.DOI
  4. [4]G. Guarino, A. Milazzo, A discontinuous Galerkin formulation for nonlinear analysis of multilayered shells refined theories, International Journal of Mechanical Sciences 255 (2023) 108426. doi:10.1016/j.ijmecsci.2023.108426.DOI
  5. [5]O. Munthe, H. P. Langtangen, Finite elements and object-oriented implementation techniques in computational fluid dynamics, Computer Methods in Applied Mechanics and Engineering 190 (8) (2000) 865–888. doi:10.1016/S0045-7825(99)00450-8.
  6. [6]Y. Bazilevs, V. M. Calo, J. A. Cottrell, T. J. R. Hughes, A. Reali, G. Scovazzi, Variational multiscale residual-based turbulence modeling for large eddy simulation of incompressible flows, Computer Methods in Applied Mechanics and Engineering 197 (1) (2007) 173–201. doi:10.1016/j.cma.2007.07.016.DOI
  7. [7]A. A. Hamada, M. Fürth, Modeling and Analysis of the Leading-Edge Vortex on Flapping Foil Turbines in Swing-Arm Mode, Journal of Fluids Engineering 145 (061105) (Mar. 2023). doi:10.1115/1.4057035.DOI
  8. [8]M. Merkel, P. Gangl, S. Schöps, Shape Optimization of Rotating Electric Machines Using Isogeometric Analysis, IEEE Transactions on Energy Conversion 36 (4) (2021) 2683–2690. doi:10.1109/TEC.2021.3061271.DOI
  9. [9]B. Kapidani, M. Merkel, S. Schöps, R. Vázquez, Tree-cotree decomposition of isogeometric mortared spaces in H(curl) on multi-patch domains, Computer Methods in Applied Mechanics and Engineering 395 (2022) 114949. doi:10.1016/j.cma.2022.114949.DOI
  10. [10]Z. Bontinck, J. Corno, S. Schöps, H. De Gersem, Isogeometric analysis and harmonic stator–rotor coupling for simulating electric machines, Computer Methods in Applied Mechanics and Engineering 334 (2018) 40–55. doi:10.1016/j.cma.2018.01.047.DOI
  11. [11]T. Hughes, J. Cottrell, Y. Bazilevs, Isogeometric analysis: CAD, finite elements, NURBS, exact geometry and mesh refinement, Computer Methods in Applied Mechanics and Engineering 194 (39-41) (2005) 4135–4195. doi:10.1016/j.cma.2004.10.008.DOI
  12. [12]J. A. Cottrell, T. J. R. Hughes, Y. Bazilevs, Isogeometric Analysis: Toward Integration of CAD and FEA, Wiley, Chichester, West Sussex, U.K Hoboken, NJ, 2009. doi:10.1002/9780470749081.DOI
  13. [13]L. Piegl, W. Tiller, The NURBS Book, Monographs in Visual Communication, Springer Berlin Heidelberg, Berlin, Heidelberg, 1997. doi:10.1007/978-3-642-59223-2.DOI
  14. [14]M. Pauley, D.-M. Nguyen, D. Mayer, J. Špeh, O. Weeger, B. Jüttler, The isogeometric segmentation pipeline, in: Isogeometric analysis and applications 2014, Springer, 2015, pp. 51–72.DOI
  15. [15]J. Guo, C. Park, D. Qian, T. J. R. Hughes, W. K. Liu, Large language model-empowered next-generation computer-aided engineering, Computer Methods in Applied Mechanics and Engineering 450 (2026) 118591. doi:10.1016/j.cma.2025.118591.DOI
  16. [16]C. S. Peskin, Flow patterns around heart valves: A numerical method, Journal of Computational Physics 10 (2) (1972) 252–271. doi:10.1016/0021-9991(72)90065-4.DOI
  17. [17]S. Osher, J. A. Sethian, Fronts propagating with curvature-dependent speed: Algorithms based on Hamilton-Jacobi formulations, Journal of Computational Physics 79 (1) (1988) 12–49. doi:10.1016/0021-9991(88)90002-2.DOI
  18. [18]F. H. Harlow, J. E. Welch, Numerical Calculation of Time-Dependent Viscous Incompressible Flow of Fluid with Free Surface, The Physics of Fluids 8 (12) (1965) 2182–2189. doi:10.1063/1.1761178.DOI
  19. [19]M. Griebel, T. Dornseifer, T. Neunhoeffer, Numerical simulation in fluid dynamics: a practical introduction, SIAM, 1998.
  20. [20]C. Hirt, B. Nichols, Volume of fluid (VOF) method for the dynamics of free boundaries, Journal of Computational Physics 39 (1) (1981) 201–225. doi:10.1016/0021-9991(81)90145-5.DOI
  21. [21]J. E. Pilliod, E. G. Puckett, Second-order accurate volume-of-fluid algorithms for tracking material interfaces, Journal of Computational Physics 199 (2) (2004) 465–502. doi:10.1016/j.jcp.2003.12.023.
  22. [22]R. J. LeVeque, Cartesian grid methods for flow in irregular regions, in: Numerical Methods for Fluid Dynamics III, 1988, pp. 375–382.
  23. [23]I. Ramière, P. Angot, M. Belliard, A fictitious domain approach with spread interface for elliptic problems with general boundary conditions, Computer Methods in Applied Mechanics and Engineering 196 (4) (2007) 766–781. doi:10.1016/j.cma.2006.05.012.DOI
  24. [24]J. Bishop, Rapid stress analysis of geometrically complex domains using implicit meshing, Computational Mechanics 30 (5) (2003) 460–478. doi:10.1007/s00466-003-0424-5.DOI
  25. [25]N. Sukumar, D. L. Chopp, N. Moës, T. Belytschko, Modeling holes and inclusions by level sets in the extended finite-element method, Computer Methods in Applied Mechanics and Engineering 190 (46) (2001) 6183–6200. doi:10.1016/S0045-7825(01)00215-8.DOI
  26. [26]A. Gerstenberger, W. A. Wall, An eXtended Finite Element Method/Lagrange multiplier based approach for fluid–structure interaction, Computer Methods in Applied Mechanics and Engineering 197 (19) (2008) 1699–1714. doi:10.1016/j.cma.2007.07.002.DOI
  27. [27]R. Becker, E. Burman, P. Hansbo, A hierarchical NXFEM for fictitious domain simulations, International Journal for Numerical Methods in Engineering 86 (4-5) (2011) 549–559. doi:10.1002/nme.3093.DOI
  28. [28]M. Schmidt, L. Noël, K. Doble, J. A. Evans, K. Maute, Extended isogeometric analysis of multi-material and multi-physics problems using hierarchical B-splines, Computational Mechanics 71 (6) (2023) 1179–1203. doi:10.1007/s00466-023-02306-x.DOI
  29. [29]E. Burman, S. Claus, P. Hansbo, M. G. Larson, A. Massing, CutFEM: Discretizing geometry and partial differential equations, International Journal for Numerical Methods in Engineering 104 (7) (2015) 472–501. doi:10.1002/nme.4823.DOI
  30. [30]J. Parvizian, A. Düster, E. Rank, Finite cell method: h- and p-extension for embedded domain problems in solid mechanics, Computational Mechanics 41 (1) (2007) 121–133. doi:10.1007/s00466-007-0173-y.DOI
  31. [31]M. Wiesheu, T. Komann, M. Merkel, S. Schöps, S. Ulbrich, I. Cortes Garcia, Combined parameter and shape optimization of electric machines with isogeometric analysis, Optimization and Engineering 26 (2) (2025) 1011–1038. doi:10.1007/s11081-024-09925-0.DOI
  32. [32]T. Komann, M. Wiesheu, S. Ulbrich, S. Schöps, Robust Design Optimization of Electric Machines with Isogeometric Analysis, Mathematics 12 (9) (Apr. 2024). doi:10.3390/math12091299.DOI
  33. [33]J. Nitsche, Über ein Variationsprinzip zur Lösung von Dirichlet-Problemen bei Verwendung von Teilräumen, die keinen Randbedingungen unterworfen sind, Abhandlungen aus dem Mathematischen Seminar der Universität Hamburg 36 (1) (1971) 9–15. doi:10.1007/BF02995904.DOI
  34. [34]V. P. Nguyen, P. Kerfriden, M. Brino, S. P. A. Bordas, E. Bonisoli, Nitsche’s method for two and three dimensional NURBS patch coupling, Computational Mechanics 53 (6) (2014) 1163–1182. doi:10.1007/s00466-013-0955-3.DOI
  35. [35]A. Razek, Y. Maday, A. Buffa, L. Santandrea, E. Bouillault, F. Rapetti, Calculation of eddy currents with edge elements on non-matching grids in moving structures, IEEE Transactions on Magnetics 36 (4) (2000) 1351–1355. doi:10.1109/20.877689.DOI
  36. [36]C. Bernardi, Y. Maday, A. Patera, A new nonconforming approach to domain decomposition: the Mortar element method, Brezis, H.(ed.) et al., Nonlinear partial differential equations and their applications. Collège de France Seminar, volume XI. Lectures presented at the weekly seminar on applied mathematics, Paris, France, (1989).
  37. [37]C. Bernardi, A new nonconforming approach to domain decomposition: ther mortar element method, Nonlinear Partial Differential Equations and their Applications (1994) 1351.
  38. [38]F. B. Belgacem, A. Buffa, Y. Maday, The mortar method for the Maxwell’s equations in 3D, Comptes Rendus de l’Académie des Sciences - Series I - Mathematics 329 (10) (1999) 903–908. doi:10.1016/S0764-4442(00)87497-2.
  39. [39]F. Ben Belgacem, A. Buffa, Y. Maday, The Mortar Finite Element Method for 3D Maxwell Equations: First Results, SIAM Journal on Numerical Analysis 39 (3) (2001) 880–901. doi:10.1137/S0036142999357968.
  40. [40]F. Rapetti, An overlapping mortar element approach to coupled magneto-mechanical problems, Mathematics and Computers in Simulation 80 (8) (2010) 1647–1656. doi:10.1016/j.matcom.2009.02.008.
  41. [41]P. Alotto, A. Bertoni, I. Perugia, D. Schotzau, Efficient use of the local discontinuous Galerkin method for meshes sliding on a circular boundary, IEEE Transactions on Magnetics 38 (2) (2002) 405–408. doi:10.1109/20.996108.DOI
  42. [42]E. Lange, F. Henrotte, K. Hameyer, A Variational Formulation for Nonconforming Sliding Interfaces in Finite Element Analysis of Electric Machines, IEEE Transactions on Magnetics 46 (8) (2010) 2755–2758. doi:10.1109/TMAG.2010.2043075.
  43. [43]E. Lange, F. Henrotte, K. Hameyer, Biorthogonal Shape Functions for Nonconforming Sliding Interfaces in 3-D Electrical Machine FE Models With Motion, IEEE Transactions on Magnetics 48 (2) (2012) 855–858. doi:10.1109/TMAG.2011.2173165.
  44. [44]S. Boehmer, E. Lange, K. Hameyer, Non-Conforming Sliding Interfaces for Relative Motion in 3D Finite Element Analysis of Electrical Machines by Magnetic Scalar Potential Formulation Without Cuts, IEEE Transactions on Magnetics 49 (5) (2013) 1833–1836. doi:10.1109/TMAG.2013.2242051.
  45. [45]F. Boy, H. Hetzler, A Non-Conformal Mapping to Avoid Mesh Deformation or Remeshing in 2D FEM-Simulation of Electrical Machines with Rotor Eccentricity, in: 2018 XIII International Conference on Electrical Machines (ICEM), IEEE, Alexandroupoli, 2018, pp. 506–512. doi:10.1109/ICELMACH.2018.8506751.DOI
  46. [46]P. Desenfans, E. Vancayseele, J. Bundschuh, H. De G ersem, Z. Gong, D. Vanoost, K. Gryllias, D. Pissoort, A Boundary-Preserving Non-Conformal Mapping for Radial Rotor Eccentricity in Finite Element Simulations of Electrical Machines, IEEE Transactions on Magnetics 60 (8) (2024) 1–8. doi:10.1109/TMAG.2024.3420813.DOI
  47. [47]S.-U. Zhang, A. V. Kumar, Magnetostatic Analysis Using Implicit Boundary Finite-Element Method, IEEE Transactions on Magnetics 46 (5) (2010) 1159–1166. doi:10.1109/TMAG.2009.2039940.DOI
  48. [48]H. De G ersem, T. Weiland, Harmonic weighting functions at the sliding interface of a finite-element machine model incorporating angular displacement, IEEE Transactions on Magnetics 40 (2) (2004) 545–548. doi:10.1109/TMAG.2004.824616.
  49. [49]X. Wei, B. Marussig, P. Antolin, A. Buffa, Immersed boundary-conformal isogeometric method for linear elliptic problems, Computational Mechanics 68 (6) (2021) 1385–1405. doi:10.1007/s00466-021-02074-6.DOI
  50. [50]G. Guarino, A. Milazzo, A. Buffa, P. Antolin, The Immersed Boundary Conformal Method for Kirchhoff–Love and Reissner–Mindlin shells, Computer Methods in Applied Mechanics and Engineering 432 (2024) 117407. doi:10.1016/j.cma.2024.117407.DOI
  51. [51]E. Lapina, P. Oumaziz, R. Bouclier, Immersed boundary-conformal isogeometric LaTIn method for multiple non-linear interfaces, Engineering with Computers (Mar. 2024). doi:10.1007/s00366-024-01946-8.DOI
  52. [52]Y. T. Elbadry, P. Antolín, O. Weeger, Immersed isogeometric analysis with boundary conformal quadrature for finite deformation elasticity, Archive of Applied Mechanics 95 (9) (2025) 219. doi:10.1007/s00419-025-02924-2.DOI
  53. [53]D. Vanoost, H. De Gersem, J. Peuteman, G. Gielen, D. Pissoort, Finite-element discretisation of the eddy-current term in a 2D solver for radially symmetric models, International Journal of Numerical Modelling: Electronic Networks, Devices and Fields 27 (3) (2014) 505–516. doi:10.1002/jnm.1955.DOI
  54. [54]A. Buffa, H. Ammari, J. C. Nédélec, A Justification of Eddy Currents Model for the Maxwell Equations, SIAM Journal on Applied Mathematics 60 (5) (2000) 1805–1823. doi:10.1137/S00361399998348979.DOI
  55. [55]D. J. Griffiths, Introduction to Electrodynamics, fourth edition Edition, Pearson, Boston, 2013.
  56. [56]S. J. Salon, Finite element analysis of electrical machines, Vol. 101, Kluwer academic publishers Boston, 1995.DOI
  57. [57]I. Babuska, The Finite Element Method with Penalty, Mathematics of Computation 27 (1973) 221–228. doi:10.1090/S0025-5718-1973-0351118-5.DOI
  58. [58]I. Babuška, The finite element method with Lagrangian multipliers, Numerische Mathematik 20 (3) (1973) 179–192. doi:10.1007/BF01436561.DOI
  59. [59]J. D. Jackson, Classical electrodynamics, John Wiley & Sons, 2012.
  60. [60]T. J. R. Hughes, The Finite Element Method: Linear Static and Dynamic Finite Element Analysis, Dover Publications, Mineola, NY, 2000.
  61. [61]P. Antolin, A. Buffa, R. Puppi, X. Wei, Overlapping Multipatch Isogeometric Method with Minimal Stabilization, SIAM Journal on Scientific Computing 43 (1) (2021) A330–A354. doi:10.1137/19M1306750.DOI
  62. [62]A. Johansson, B. Kehlet, M. G. Larson, A. Logg, Multimesh finite element methods: Solving PDEs on multiple intersecting meshes, Computer Methods in Applied Mechanics and Engineering 343 (2019) 672–689.
  63. [63]P. Antolin, A. Buffa, M. Martinelli, Isogeometric Analysis on V-reps: First results, Computer Methods in Applied Mechanics and Engineering 355 (2019) 976–1002. doi:10.1016/j.cma.2019.07.015.DOI
  64. [64]G. Elber, The irit geometric modeling environment, https://csaws.cs.technion.ac.il/~gershon/irit (2023).
  65. [65]Open CASCADE SAS, Open cascade technology, https://www.opencascade.com, version 7.7.0 (2022).
  66. [66]X. Wei, R. Puppi, P. Antolin, A. Buffa, Stabilized isogeometric formulation of the stokes problem on overlapping patches, Computer Methods in Applied Mechanics and Engineering 417 (2023) 116477.arxiv.org/abs/2305.20021
  67. [67]M. G. Larson, A. Logg, C. Lundholm, Space-time CutFEM on overlapping meshes I: Simple continuous mesh motion, Numerische Mathematik 156 (3) (2024) 1015–1054. doi:10.1007/s00211-024-01417-8.DOI
  68. [68]M. G. Larson, C. Lundholm, Space-time CutFEM on overlapping meshes II: Simple discontinuous mesh evolution, Numerische Mathematik 156 (3) (2024) 1055–1083. doi:10.1007/s00211-024-01413-y.DOI
  69. [69]F. de Prenter, C. V. Verhoosel, E. H. van Brummelen, M. G. Larson, S. Badia, Stability and Conditioning of Immersed Finite Element Methods: Analysis and Remedies, Archives of Computational Methods in Engineering 30 (6) (2023) 3617–3656. doi:10.1007/s11831-023-09913-0.DOI
  70. [70]A. Buffa, R. Puppi, R. Vázquez, A Minimal Stabilization Procedure for Isogeometric Methods on Trimmed Geometries, SIAM Journal on Numerical Analysis 58 (5) (2020) 2711–2735. doi:10.1137/19M1244718.DOI
  71. [71]F. De Prenter, C. Lehrenfeld, A. Massing, A note on the stability parameter in Nitsche’s method for unfitted boundary value problems, Computers & Mathematics with Applications 75 (12) (2018) 4322–4336. doi:10.1016/j.camwa.2018.03.032.DOI
  72. [72]K. Höllig, U. Reif, J. Wipper, Weighted Extended B-Spline Approximation of Dirichlet Problems, SIAM Journal on Numerical Analysis 39 (2) (2001) 442–462. doi:10.1137/S0036142900373208.DOI
  73. [73]D. Elfverson, M. G. Larson, K. Larsson, A new least squares stabilized Nitsche method for cut isogeometric analysis, Computer Methods in Applied Mechanics and Engineering 349 (2019) 1–16. doi:10.1016/j.cma.2019.02.011.DOI
  74. [74]T. Jonsson, M. G. Larson, K. Larsson, Graded parametric CutFEM and CutIGA for elliptic boundary value problems in domains with corners, Computer Methods in Applied Mechanics and Engineering 354 (2019) 331–350. doi:10.1016/j.cma.2019.05.024.DOI
  75. [75]B. Marussig, J. Zechner, G. Beer, T.-P. Fries, Stable isogeometric analysis of trimmed geometries, Computer Methods in Applied Mechanics and Engineering 316 (2017) 497–521. doi:10.1016/j.cma.2016.07.040.DOI
  76. [76]B. Marussig, R. Hiemstra, T. J. R. Hughes, Improved conditioning of isogeometric analysis matrices for trimmed geometries, Computer Methods in Applied Mechanics and Engineering 334 (2018) 79–110. doi:10.1016/j.cma.2018.01.052.DOI
  77. [77]C. Lang, D. Makhija, A. Doostan, K. Maute, A simple and efficient preconditioning scheme for heaviside enriched XFEM, Computational Mechanics 54 (5) (2014) 1357–1374. doi:10.1007/s00466-014-1063-8.DOI
  78. [78]C. Lehrenfeld, A. Reusken, Optimal preconditioners for Nitsche-XFEM discretizations of interface problems, Numerische Mathematik 135 (2) (2017) 313–332. doi:10.1007/s00211-016-0801-6.DOI
  79. [79]F. de Prenter, C. V. Verhoosel, G. J. van Zwieten, E. H. van Brummelen, Condition number analysis and preconditioning of the finite cell method, Computer Methods in Applied Mechanics and Engineering 316 (2017) 297–327. doi:10.1016/j.cma.2016.07.006.DOI
  80. [80]F. de Prenter, C. V. Verhoosel, E. H. van Brummelen, J. A. Evans, C. Messe, J. Benzaken, K. Maute, Multigrid solvers for immersed finite element methods and immersed isogeometric analysis, Computational Mechanics 65 (3) (2020) 807–838. doi:10.1007/s00466-019-01796-y.DOI
  81. [81]R. Vázquez, A new design for the implementation of isogeometric analysis in Octave and Matlab: GeoPDEs 3.0, Computers & Mathematics with Applications 72 (3) (2016) 523–554. doi:10.1016/j.camwa.2016.05.010.DOI
  82. [82]R. Vázquez, GeoPDEs: a package for Isogeometric Analysis in Matlab and Octave, https://rafavzqz.github.io/geopdes, version 3.4.2 (2024).
  83. [83]H. Houssein, Symmetric Formulation for Non-Conforming Meshes in Magnetostatic Problems, IEEE Transactions on Magnetics 60 (4) (2024) 1–7. doi:10.1109/TMAG.2024.3358546.DOI
  84. [84]A. Kitanovski, Energy Applications of Magnetocaloric Materials, Advanced Energy Materials 10 (10) (2020) 1903741. doi:10.1002/aenm.201903741.DOI
  85. [85]M. Wiesheu, M. Merkel, T. Sittig, D. Benke, M. Fries, S. Schöps, O. Weeger, I. C. Garcia, How to Build the Optimal Magnet Assembly for Magnetocaloric Cooling: Structural Optimization with Isogeometric Analysis, International Journal of Refrigeration 152 (2023) 62–73. doi:10.1016/j.ijrefrig.2023.04.014.DOI
  86. [86]R. Bjork, C. R. H. Bahl, A. Smith, N. Pryds, Improving Magnet Designs With High and Low Field Regions, IEEE Transactions on Magnetics 47 (6) (2011) 1687–1692. doi:10.1109/TMAG.2011.2114360.DOI
  87. [87]C. Giannelli, B. Jüttler, H. Speleers, THB-splines: The truncated basis for hierarchical splines, Computer Aided Geometric Design 29 (7) (2012) 485–498. doi:10.1016/j.cagd.2012.03.025.DOI
  88. [88]A. Grendas, M. Wiesheu, S. Schöps, B. Marussig, Adaptive Isogeometric Analysis With THB-Splines and Multi-Level Bézier Extraction for Coupled Magnetostatics, IEEE Transactions on Magnetics 62 (1) (2026) 1–11. doi:10.1109/TMAG.2025.3640276.DOI
  89. [89]P. Gangl, N. Krenn, H. De Gersem, Multi-material topology optimization of electric machines under maximum temperature and stress constraints, Applied Mathematical Modelling 150 (2026) 116347. doi:10.1016/j.apm.2025.116347.DOI
  90. [90]OpenAI, ChatGPT, https://chat.openai.com (2026).

Paper details

vlnc:W_gapfsdf2
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 8

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

ChatGPT

View statement
Looking for related papers…
Contents