betaFeedback

WorksMathematics

A time-domain adversarial example demonstrating the failure of H1H^1 elements in aa-ϕϕ electromagnetic simulation

Abstract

Frequently, H1H^1 continuous finite elements are used in computational electromagnetics despite it being widely known that such an interpolation is not guaranteed to furnish reliable approximations of solutions to general initial/boundary value problems. It is mathematically more natural to use H(Curl)H(\text{Curl}) elements (edge elements) than H1H^1 elements, since the former form part of a discrete complex that respects the differential-geometric structure of Maxwell's equations while the latter do not. However, most work discussing the failure of H1H^1 elements for electromagnetics is esoteric and concrete examples demonstrating the clear advantage of edge elements are generally restricted to quasi-static, frequency-domain, or eigenvalue problems. In this work, we summarize essential facts regarding H1H^1 and H(Curl)H(\text{Curl}) elements and present a physically feasible yet intentionally adversarial time-domain eddy-current example demonstrating the failure of gauge-penalized H1H^1 elements in an aa-ϕϕ formulation compared to gauge-free mixed scalar/edge H1/H(Curl)H^1/H(\text{Curl}) elements. The results indicate that edge elements can be used to furnish reasonable results for general loading conditions, whereas penalized H1H^1 elements fail (e.g., converge to an incorrect state) even with mesh refinement.

Introduction

Although the advantages and useful properties of H(Curl)H(\mathrm{Curl}) edge elements are well known and have been thoroughly studied (see, e.g., (see, e.g., Zaglmayr, 2006)), many works continue to use H1H^1 vector-valued elements in time-domain, aa-ϕ\phi electromagnetics applications (see, e.g., (see, e.g., Moreno-Mateos et al. (2023), Stewart and Anand (2026), Dorn et al. (2021), and Lucarini et al. (2022))). It is not unreasonable to assume that this continued prevalence may owe itself to the complexity of the analytic explanations of the failure of H1H^1 elements compared to H(Curl)H(\mathrm{Curl}) edge elements and further to the benign nature of many benchmark problems, which fail to clearly elucidate the underlying issues. Additionally, while there exist numerous examples demonstrating that incorrect eigenvalues are obtained with H1H^1 elements (see, e.g., Boffi et al. [5] and Cendes [11], the time domain eddy current examples present in the literature seem to focus on the comparative accuracy of global responses (see, e.g., [21, 19, 25]). Even when edge elements are deemed superior in such cases, the examples do not elucidate the fact that structural issues plaguing H1H^1 elements can lead to obviously spurious solutions in the time domain as well. Finally, many texts seemingly frame the choice of vector function spaces as an opportunity for optional improvement rather than as a necessary means of avoiding issues caused by an inherent analytical deficiency. Given this setting, it is understandable why many, especially those who are new to this particular class of problems or are approaching the field from a multiphysics perspective, may incorrectly view an H1H^1 interpolation as a harmless simplification.

The focus of this work is twofold. First, we aim to provide a concise explanation of the most important analytical reasons that H1H^1 elements should be avoided in favor of H(Curl⁡)H(\operatorname{Curl}) edge elements in any aa-ϕ\phi formulation of electromagnetics. Second, we provide a concrete example in which H1H^1 elements fail to furnish a reliable approximation to the solution of a boundary value problem compared to H(Curl⁡)H(\operatorname{Curl}) elements, thereby making the issue concrete even to readers for whom the full analytical explanation may be difficult to parse. In the example, the quality of the solution is obviously very poor when fully H1H^1 elements are used, an issue that persists with refinement, while H(Curl⁡)H(\operatorname{Curl}) elements converge towards the correct solution. Through these explanations and the associated demonstrations, we illustrate that H1H^1 elements are structurally deficient in their ability to approximate solutions to problems in aa-ϕ\phi electromagnetics formulations compared to H(Curl⁡)H(\operatorname{Curl}) elements.

Theoretical Background

Maxwell's equations

The electromagnetic balance laws are widely summarized in many works (see, e.g., Kovetz [20], among others). In media, Maxwell’s equations read as

Gauss Law:∮∂Bd⋅n da=∫Bq dv(1)\begin{aligned} \text{Gauss Law:}\quad&\oint_{\partial B} \mathbf{d}\cdot\mathbf{n}\,\mathrm{d}a = \int_B q\,\mathrm{d}v \tag*{(1)} \end{aligned}
Maxwell-Ampeˋre Law:∮∂Ah⋅dx=∫Ai⋅n da+∂∂t∫Ad⋅n da(2)\begin{aligned} \text{Maxwell-Ampère Law:}\quad&\oint_{\partial A} \mathbf{h}\cdot\mathrm{d}\mathbf{x} = \int_A \mathbf{i}\cdot\mathbf{n}\,\mathrm{d}a + \frac{\partial}{\partial t}\int_A \mathbf{d}\cdot\mathbf{n}\,\mathrm{d}a \tag*{(2)} \end{aligned}
Faraday Law:∮∂Ae⋅dx=−∂∂t∫Ab⋅n da(3)\begin{aligned} \text{Faraday Law:}\quad&\oint_{\partial A} \mathbf{e}\cdot\mathrm{d}\mathbf{x} = -\frac{\partial}{\partial t}\int_A \mathbf{b}\cdot\mathbf{n}\,\mathrm{d}a \tag*{(3)} \end{aligned}
Monopole Law:∮∂Bb⋅n da=0(4)\begin{aligned} \text{Monopole Law:}\quad&\oint_{\partial B} \mathbf{b}\cdot\mathbf{n}\,\mathrm{d}a = 0 \tag*{(4)} \end{aligned}

where d\mathbf{d} is the electric displacement field1 [C/m2][\mathrm{C}/\mathrm{m}^2], qq is the macroscopic free charge density [C/m3][\mathrm{C}/\mathrm{m}^3], h\mathbf{h} is the magnetic field2 [A/m][\mathrm{A}/\mathrm{m}], i\mathbf{i} is the free current density [A/m2][\mathrm{A}/\mathrm{m}^2], e\mathbf{e} is the electric field [V/m][\mathrm{V}/\mathrm{m}], and b\mathbf{b} is the magnetic flux density3 [T][\mathrm{T}]. Here, B\mathcal{B} and A\mathcal{A} are arbitrary volumes and surfaces with unit normal vector fields n\mathbf{n}, taken outward on ∂B\partial B. It is worth noting that much of the difficulty associated with the interpolation of electromagnetic quantities is that as expressed in (1)–(4), the fields are represented by their vector proxies, whereas in their “true definition” they are differential forms which respect a strict differential-geometric structure (see, e.g., Bossavit [6]). The existence of the vector proxies is a consequence of the special structure of R4\mathbb{R}^{4} (space-time),4 and while the vector proxies make electromagnetics more accessible, they hide essential and important geometric structure of the governing equations and the fields.

In a stationary linear isotropic dielectric or magnetic material and a linear isotropic ohmic conductor, to close the system of equations, (1)–(4) are supplemented with the following constitutive equations:

d=ϵ0(1+χ)e,h=μ0−1(1−χm)b,i=κe.(5)\mathbf{d}=\epsilon_0(1+\chi)\mathbf{e}, \qquad\mathbf{h}=\mu_0^{-1}(1-\chi_m)\mathbf{b}, \qquad\mathbf{i}=\kappa\mathbf{e}. \tag*{(5)}

These are simplifications of more general models derivable from the electromagnetic duality relations and thermodynamic arguments (see e.g., Carlson and Govindjee [10]). Here ϵ0\epsilon_0, μ0\mu_0, κ\kappa, χ\chi, and χm\chi_m are the (electric) permittivity of free space ≈8.854×10−12\approx8.854 \times10^{-12} [F/m], (magnetic) permeability of free space ≈4π×10−7\approx4\pi\times10^{-7} [T⋅\cdotm/A], electric conductivity [S/m], dielectric susceptibility, and magnetic susceptibility, respectively. Also worth noting is that (μ0ϵ0)−1=c2(\mu_0\epsilon_0)^{-1}=c^2, the speed of light squared.

Time domain aa-ϕ\phi formulation

On a contractible space (e.g., Rn\mathbb{R}^{n}), the Poincaré lemma (see, e.g., Lee [22]) implies that Faraday’s and the Monopole laws (3) and (4) are automatically satisfied by introducing potentials a\mathbf{a} [T⋅\cdotm] and ϕ\phi [V] in terms of which one declares that

b=∇×a,e=−a,t−∇ϕ.(6)\mathbf{b}=\nabla\times\mathbf{a}, \qquad\mathbf{e}=-\mathbf{a}_{,t}-\nabla\phi. \tag*{(6)}

Therefore, Maxwell’s equations can be reduced locally to

div⁡[ϵ0(1+χ)(a,t+∇ϕ)]+q=0(7)\operatorname{div}\left[\epsilon_0(1+\chi)(\mathbf{a}_{,t}+\nabla\phi)\right]+q=0 \tag*{(7)}
μ0−1∇×[(1−χm)∇×a]+κ(a,t+∇ϕ)+ϵ0[(1+χ)(a,t+∇ϕ)],t=0(8)\mu_0^{-1}\nabla\times\left[(1-\chi_m)\nabla\times\mathbf{a}\right]+\kappa(\mathbf{a}_{,t}+\nabla\phi)+\epsilon_0\left[(1+\chi)(\mathbf{a}_{,t}+\nabla\phi)\right]_{,t}=0 \tag*{(8)}

modulo a to-be-discussed gauge constraint, say, the Coulomb or Lorentz gauge (Jackson [16]). In many time domain applications, the frequency ω\omega of electromagnetic loading is such that ωϵ0≪κ\omega\epsilon_0\ll\kappa, making the third term in the Maxwell-Ampère law negligible in comparison to the second term. Further, materials can be assumed to be electro-neutral at the macroscopic scale so that q=0q=0 in most cases. Thus, in practical applications it is often the case that the equations

div⁡[ϵ0(1+χ)(a,t+∇ϕ)]=0(9)\operatorname{div}\left[\epsilon_0(1+\chi)(\mathbf{a}_{,t}+\nabla\phi)\right]=0 \tag*{(9)}
μ0−1∇×[(1−χm)∇×a]+κ(a,t+∇ϕ)=0(10)\mu_0^{-1}\nabla\times\left[(1-\chi_m)\nabla\times\mathbf{a}\right]+\kappa(\mathbf{a}_{,t}+\nabla\phi)=0 \tag*{(10)}

determine the electromagnetic fields to good accuracy. It is worth remarking that, in materials with piecewise constant properties, (9) is locally equivalent to the charge conservation equation

div⁡[κ(a,t+∇ϕ)]=0.(11)\operatorname{div}\left[\kappa\left(\boldsymbol{a}_{,t}+\nabla\phi\right)\right]=0. \tag*{(11)}

The choice of which equation to solve, (9) or (11), typically depends on whether the materials of interest are conductors or dielectrics; for example, in an eddy current application (11) would be preferred.

In this work, only the aa-ϕ\phi formulation will be considered given its ability to be extended to a very general class of problems. Moreover, an aa-ϕ\phi formulation is often necessary in multiphysics applications involving electrochemistry, where the electromagnetic potentials become physically important in constitutive relations rather than just mathematically convenient (see, e.g., [10]).

H1H^{1} and H(Curl)H(\mathrm{Curl}) elements: The discrete de Rham complex

The potential formulation of Maxwell’s equations is valid in large part thanks to the exactness of the de Rham complex over R3\mathbb{R}^{3} or any contractible smooth submanifold (see, e.g., [22 Chap. 17]). Letting Πi\Pi_i denote the projections onto the finite element interpolation spaces used, one should ensure that the diagram

H1→∇H(Curl)→∇×H(Div)→div⁡L2Πϕ↓Πa↓Π2↓Π3↓Sϕ→∇dNa→∇d×Wb→div⁡dDq\begin{CD} H^{1} @>{\nabla}>> H(\mathrm{Curl}) @>{\nabla\times}>> H(\mathrm{Div}) @>{\operatorname{div}}>> L^{2} \\ @V{\Pi_{\phi}}VV @V{\Pi_{a}}VV @V{\Pi_{2}}VV @V{\Pi_{3}}VV \\ S_{\phi} @>{\nabla_d}>> N_a @>{\nabla_d\times}>> W_b @>{\operatorname{div}_d}>> D_q \end{CD}

commutes in order that approximated solutions satisfy the geometric structure of Maxwell’s equations (see, e.g., [15]). Here, SϕS_{\phi} denotes the interpolation space for ϕ\phi, NaN_a denotes the interpolation space for aa, WbW_b denotes an interpolation space for 2-form proxies like bb, DqD_q denotes an interpolation space for 3-form proxies like qq, and the dd subscript denotes the action of the discrete differential operator.5 For a given domain D\mathcal{D}, these spaces are subspaces of H1={v∈L2(D):∇v∈L2(D)3}H^{1}=\{v\in L^{2}(\mathcal{D}):\nabla v\in L^{2}(\mathcal{D})^{3}\}, H(Curl)={v∈L2(D)3:∇×v∈L2(D)3}H(\mathrm{Curl})=\{\boldsymbol{v}\in L^{2}(\mathcal{D})^{3}:\nabla\times\boldsymbol{v}\in L^{2}(\mathcal{D})^{3}\}, H(Div)={v∈L2(D)3:div⁡v∈L2(D)}H(\mathrm{Div})=\{\boldsymbol{v}\in L^{2}(\mathcal{D})^{3}:\operatorname{div}\boldsymbol{v}\in L^{2}(\mathcal{D})\}, and L2={v measurable:∫D∣v∣2 dx<∞}L^{2}=\{v\text{ measurable}:\int_{\mathcal{D}}|v|^{2}\,dx<\infty\}, respectively, with derivatives understood weakly.

By commutation we mean the interpolation of the gradient of a scalar function should always be equivalent to the discrete gradient of the interpolation of that scalar function, and so on. Interpreted physically, the issue of interpolation is not merely whether the polynomial order is high enough, but whether the discrete spaces preserve the differential relationships among potentials, fields, fluxes, and charges.

A primary advantage of H(Curl)H(\mathrm{Curl}) edge elements is their role in such a discrete complex, enabling good approximations of boundary value problem solutions to the order of approximation. On top of this, assignment of boundary conditions is far simpler and normal continuity of the time derivative of the vector potential need not be maintained; such discontinuities are often necessary for consistent solutions (see, e.g., Jin, 2014). On the other hand, using H1H^1 elements to approximate, for example, vector (1-form) objects like a\mathbf{a} is problematic in general given that these elements are not a part of any sequence satisfying this projection property. Therefore, satisfying Maxwell’s equations in a weak, discrete setting with H1H^1 elements does not imply that the finite element solution is an approximation of the continuous, analytical problem in general, especially when the true solution is not H1H^1 continuous.

Numerical solution strategies

Weak form equations

The equations to be solved in the present context over an approximation domain D\mathcal{D} can be stated in weak form as

∫Dκ(a,t+∇ϕ)⋅∇δϕ dv+∫∂Dδϕ i⋅n da=0,(12)\int_{\mathcal{D}} \kappa(\mathbf{a}_{,t}+\nabla\phi)\cdot\nabla\delta\phi\,dv+\int_{\partial\mathcal{D}}\delta\phi\,\mathbf{i}\cdot\mathbf{n}\,da=0, \tag*{(12)}
∫D(1−χm)[∇×a]⋅[∇×δa]+μ0κ(a,t+∇ϕ)⋅δa dv+∫∂D(1−χm)(n×b)⋅δa da=0,(13)\int_{\mathcal{D}}(1-\chi_m)[\nabla\times\mathbf{a}]\cdot[\nabla\times\delta\mathbf{a}]+\mu_0\kappa(\mathbf{a}_{,t}+\nabla\phi)\cdot\delta\mathbf{a}\,dv+\int_{\partial\mathcal{D}}(1-\chi_m)(\mathbf{n}\times\mathbf{b})\cdot\delta\mathbf{a}\,da=0, \tag*{(13)}

where δϕ\delta\phi and δa\delta\mathbf{a} are test functions. In general, the solution to this set of equations requires an additional specification of a gauge controlling the gradient null-space of a\mathbf{a} and the time dependent null-space of ϕ\phi. Heuristically, these are the “rigid body” modes; if a^\hat{\mathbf{a}} and ϕ^\hat{\phi} are solutions, so are a^+∇f\hat{\mathbf{a}}+\nabla f and ϕ^−f,t\hat{\phi}-f_{,t} for any scalar field ff. A common and simple choice is the Coulomb gauge condition, div⁡a=0\operatorname{div}\mathbf{a}=0 [16 §6.3].

Edge element formulations

In this work, mixed edge scalar H1/H(Curl⁡)H^1/H(\operatorname{Curl}) elements, H1H^1 for ϕ\phi and H(Curl⁡)H(\operatorname{Curl}) for a\mathbf{a}, are used to interpolate solutions to boundary value problems. The core advantage of edge elements is that the associated interpolation space NaN_a is a part of a discrete exact complex. For the transient problems presently considered, the vector-potential block of the time-discretized form of Eq. (13) is coercive in the H(Curl⁡)H(\operatorname{Curl}) norm,6 and no additional algebraic gauge constraint need be imposed. The commuting discretization preserves the required compatibility structure at the discrete level, allowing the gauge associated with compatible initial data to evolve consistently while leaving the computed electromagnetic fields invariant. This is analogous to the way that the rigid body motion in a dynamic elasticity problem evolves in time; the numerically imposed gauge evolves consistently in time based on the gauge of the initial conditions.

The mixed edge scalar elements implemented in this study are 20 and 32 node hexahedral elements, as depicted in Figure 1. The 20 node elements consist of eight corner nodes associated with first order, trilinear degrees of freedom used to interpolate the scalar potential ϕ\phi and a single node on each edge associated with zeroth-order complete Nédélec edge degrees of freedom used to interpolate the magnetic vector potential a\boldsymbol{a}. These two spaces form a part of a discrete exact complex, meaning that ∇Sϕ=ker⁡(∇×∣Na)\nabla S_{\phi} = \ker(\nabla\times\vert_{N_a}), mirroring the exact continuous complex.

Depictions of the 20 (left) and 32 (right) node elements utilized in this work, with the typical 8 node hexahedral reference element shown in the middle

Figure 1. Depictions of the 20 (left) and 32 (right) node elements utilized in this work. The typical 8 node hexahedral reference element is also shown (middle).

In the time domain, exactness of the discrete complex is not strictly necessary to furnish good solutions to boundary value problems with mixed scalar/edge interpolations so long as the exact complex associated with the lowest order space is a subcomplex of the exact complex associated with the highest order space. The commutative nature of the lowest order diagram ensures that the higher order quantity is projected consistently in the lowest order equation. Thus, any error associated with this particular form of inexactness is identical in nature to the error associated with the general interpolation of, for example, a nonlinear solution with linear elements, and is not a more fundamental structural issue leading to genuinely spurious solutions at lowest order. To demonstrate how these elements can be effective, the 32 node elements consist of the same scalar corner degrees of freedom as in the 20 node case, only now with 24 first-order complete serendipity edge degrees of freedom for the magnetic vector potential with two degrees of freedom associated with each edge (Arnold and Awanou, [2]). These serendipity H(Curl)H(\mathrm{Curl}) functions are a part of a discrete exact complex containing the complex associated with the first-order H1H^1 scalar degrees of freedom as a subcomplex. Again, although the discrete complex formed by SϕS_{\phi} and NaN_a is now not exact, the continuous geometric structure is still maintained to lowest order while simultaneously allowing for a higher order interpolation of a degree of freedom of choice: in this case, the vector potential a\boldsymbol{a}. Thus, one can achieve a higher order of accuracy with a reduced number of degrees of freedom; in this case, the first-order H(Curl)H(\mathrm{Curl}) elements have the same number of degrees of freedom as the fully H1H^1 elements.7

Fully H1H^1 element formulations

As noted, fully H1H^1 elements where each component of a\boldsymbol{a} is interpolated as an H1H^1 continuous scalar are not a part of a discrete exact complex that forms a commutative diagram with the continuous de Rham sequence. Therefore, modes that discretely satisfy (12) and (13) may not correspond to actual, continuous fields associated with the boundary and initial data. In many cases, this failure is especially evident for the gradient part of a\boldsymbol{a} in the Helmholtz decomposition. To stabilize/control these spurious modes, a penalty term is often added to (13) in H1H^1 formulations to enforce the Coulomb gauge, leading to a modified weak form:

∫D(1−χm)([∇×a]⋅[∇×δa]+λdiv⁡(a)div⁡(δa))+μ0κ(a,t+∇φ)⋅δa dv+∫∂D(1−χm)(n×b)⋅δa da=0,(14)\int_{\mathcal{D}} (1-\chi_m)\left([\nabla\times\boldsymbol{a}]\cdot[\nabla\times\delta\boldsymbol{a}]+\lambda\operatorname{div}(\boldsymbol{a})\operatorname{div}(\delta\boldsymbol{a})\right)+\mu_0\kappa(\boldsymbol{a}_{,t}+\nabla\varphi)\cdot\delta\boldsymbol{a}\,dv+\int_{\partial\mathcal{D}}(1-\chi_m)(\boldsymbol{n}\times\boldsymbol{b})\cdot\delta\boldsymbol{a}\,da=0, \tag*{(14)}

where λ\lambda is a penalty parameter often set to unity (see, e.g., Coulomb (1981), Jin (2014, §5.7.4), or Anand et al. (2026)).8 This gauge is generally necessary since the structure of the function spaces does not allow for consistent gauge evolution in time without explicit control.

It should be noted that, even if strictly enforced, this constraint is not enough to ensure that a discretely computed curl corresponds to the compatible projection of the continuous curl associated with the boundary data. To fully recover the correct geometric structure, one should also solve div⁡b=0\operatorname{div}\boldsymbol{b}=0 with a separate interpolation for b\boldsymbol{b} under the constraint b=∇×a\boldsymbol{b}=\nabla\times\boldsymbol{a}. Although seemingly redundant, strictly speaking such compatibility is necessary given that, for arbitrary a′∈H(Curl⁡)\boldsymbol{a}'\in H(\operatorname{Curl}), it may be the case that Π2(∇×a′)≠∇d×(Πaa′)\Pi_2(\nabla\times\boldsymbol{a}')\ne\nabla_d\times(\Pi_a\boldsymbol{a}'). In other words, even a strict divergence constraint on a\boldsymbol{a} cannot guarantee the absence of spurious solutions given that a discretely computed curl need not correspond to the continuous curl of interest.

The H1H^1 elements implemented in this work are 32 degree of freedom eight node trilinear hexahedral elements consisting of four degrees of freedom at each node: three to interpolate each component of the magnetic vector potential a\boldsymbol{a} and one to interpolate the electric scalar potential. This element is graphically depicted in Figure 1.

Solver technology

The linearized system of equations for problems posed by the theory is ill conditioned, even with spatially uniform conductivity values. To still provide reasonable solutions at an acceptable tolerance,9 the field-split preconditioning methodology of Carlson and Govindjee [10] is utilized in PETSc with a flexible, iterative GMRES solver [4, 3]. This solver, along with the elements discussed, is implemented in FEAP [27]. Time stepping is carried out using FEAP’s BDF2 time integrator, and all numerical integration is computed via a 3×3×33\times3\times3 Gaussian quadrature rule.

Adversarial benchmark problem

To illustrate the consequences of the limitations of the H1H^1 approximation discussed above, a bi-material square prism subjected to a linearly increasing magnetic field is considered. The geometry and loading admit an independently constructed semi-analytic reference solution, allowing the current density computed by the finite element formulations to be assessed directly. After a short startup transient, the induced current becomes independent of time, providing a simple setting in which to examine approximation error in space and time.

Geometry, material properties, and loading

The specimen occupies a square prism of dimensions 1.51.5 mm ×1.5\times1.5 mm ×0.5\times0.5 mm, centered at the origin. An aluminum alloy conducting annulus surrounds a titanium alloy core with a 0.50.5 mm ×0.5\times0.5 mm cross section extending through the full thickness of the specimen. Thus, the annular band has a thickness of 0.50.5 mm on each side. The material properties are idealized as indicated in Table 1; in particular, the conductivity contrast is prescribed to be exactly ten to one.10 Both materials are treated as non-magnetic, and perfect electrical contact is assumed at their interface.

ParameterAluminum alloy annulusTitanium alloy core
Conductivity κ\kappa [S/m]κa=2.538×107\kappa_a = 2.538 \times10^7κa/10=2.538×106\kappa_a/10 = 2.538 \times10^6
Magnetic susceptibility χm\chi_m [−-]00

Table 1. Simulation model parameters

Denote the domain by D\mathcal{D} and its exterior boundary by ∂D\partial\mathcal{D}. The imposed magnetic boundary condition and scalar potential are11

n×b=n×(βtE^3),ϕ=0on ∂D,β=1 T/s.(15)\boldsymbol{n} \times\boldsymbol{b} = \boldsymbol{n} \times(\beta t \widehat{\boldsymbol{E}}_3), \qquad\phi= 0 \quad\text{on } \partial\mathcal{D}, \qquad\beta= 1\ \mathrm{T/s}. \tag*{(15)}

The ramp begins at t=0t = 0 from an initially field-free, current-free state, with a0=0\boldsymbol{a}_0 = 0 and ϕ0=0\phi_0 = 0. In the magneto-quasi-static limit of the Maxwell–Ampère law, Eq. (2), the prescribed tangential magnetic field has zero circulation around any closed loop on an exterior face and therefore implies i⋅n=0\boldsymbol{i} \cdot\boldsymbol{n} = 0 on each such face. The geometry and boundary conditions are depicted in Figure 2.

Figure 2. Geometry and loading of the bi-material conductor. The titanium core extends through the full thickness of the aluminum annulus. On any given exterior face with normal n\mathbf{n} and at any time tt, the applied magnetic boundary condition is n×b=βt(n×E^3)\mathbf{n} \times\mathbf{b} = \beta t(\mathbf{n} \times\widehat{\mathbf{E}}_3).

The fixed scalar potential is interpreted as a prescribed electrochemical reference state. A possible physical idealization is a specimen equilibrated in a reversible redox bath, with the whole assembly placed in a nearly uniform, axially ramped solenoid b\boldsymbol{b}-field. Given the small specimen size and electric field magnitudes, it is reasonable to assume that chemical equilibrium is approximately maintained at the specimen–bath interface and to neglect the microtesla-scale self-field when identifying the prescribed boundary field with the applied field.

Remark. Finite element solutions are computed on three-dimensional meshes despite the two-dimensional nature of the eddy-current solution.

Semi-analytic reference solution

By symmetry of the geometry and loading, the steady-ramp magnetic field can be written as b=(βt−g)E^3\mathbf{b} = (\beta t - g)\widehat{\mathbf{E}}_3, where g=g(x,y)g = g(x,y) vanishes on the perimeter of the cross section. The constitutive relations in Eq. (5), together with the magneto-quasi-static Maxwell–Ampère and Faraday laws, Eqs. (2) and (3), reduce the problem to

−div⁡(κ−1∇g)=μ0β.(16)-\operatorname{div}(\kappa^{-1}\nabla g) = \mu_0\beta. \tag*{(16)}

Because both materials are non-magnetic and perfect electrical contact is assumed, gg and κ−1∇g⋅n\kappa^{-1}\nabla g \cdot\mathbf{n} are continuous across material interfaces.12 The resulting eddy current is

iref=1μ0(∂g∂xE^2−∂g∂yE^1).(17)\mathbf{i}_{\mathrm{ref}} = \frac{1}{\mu_0}\left(\frac{\partial g}{\partial x}\widehat{\mathbf{E}}_2 - \frac{\partial g}{\partial y}\widehat{\mathbf{E}}_1\right). \tag*{(17)}

Details regarding the semi-analytic solution for gg are given in Appendix § 8.

Results

In all finite element calculations, the volume elements are cubes and the time step is fixed at Δt=5 μs\Delta t = 5\,\mu\mathrm{s}. Computations are carried out on a single processor and, unless otherwise specified, the penalty parameter is set to λ=1\lambda= 1 in H1H^1 element formulations.

Solution comparisons

Figure 3 compares the computed current fields on the coarsest mesh at t=1t = 1 ms with the semi-analytic steady-ramp solution of Eq. (17). In the reference solution, the current circulates through the aluminum annulus, follows the exterior faces, and vanishes at the four outer corners of the cross section. However, the Coulomb-penalized H1H^1 solution places its largest currents at precisely these corners. Additionally, along the adjacent faces, the current vectors exhibit a pronounced component normal to the exterior boundary, in violation of the no-through-current condition i⋅n=0\boldsymbol{i} \cdot\boldsymbol{n} = 0 implied by the magnetic field boundary condition. The H(Curl)H(\mathrm{Curl}) solutions more closely reproduce both the distribution of current in the annulus and its direction near the boundary. Moreover, for both H(Curl)H(\mathrm{Curl}) formulations, the settled eddy-current profile approaches the reference profile as the mesh is refined (Figure 4). In contrast, on all three meshes the settled Coulomb-penalized H1H^1 solution remains similar to the coarse-mesh solution shown in Figure 3(b).

Figure 3. Eddy-current distributions on the coarsest mesh under a linear magnetic-field ramp. The finite element fields and the semi-analytic reference are shown with identical limits for current magnitude and a common scale for the current vectors. The finite element mesh is superimposed on the semi-analytic reference field solely to aid comparison.

Figure 4. Relative current error for the Coulomb-penalized H1H^1, the order-zero H(Curl)H(\mathrm{Curl}), and the order-one H(Curl)H(\mathrm{Curl}) element formulations. Here, hh denotes the element edge length, and the middle and right panels use semi-log axes.

Effect of mesh refinement

Figure 4 compares three meshes containing 72, 576, and 4608 volume elements, obtained by successively halving the element edge length. The relative current error,

eˉ(t)=∥ih(t)−iref∥κ−1∥iref∥κ−1,(18)\bar{e}(t) = \frac{\left\Vert\boldsymbol{i}_h(t) - \boldsymbol{i}_{\mathrm{ref}} \right\Vert_{\kappa^{-1}}}{\left\Vert\boldsymbol{i}_{\mathrm{ref}} \right\Vert_{\kappa^{-1}}}, \tag*{(18)}

is evaluated from the element fields at volume quadrature points using the resistivity-weighted norm

∥i∥κ−12=∫Dκ−1∣i∣2 dv.(19)\left\Vert\boldsymbol{i} \right\Vert_{\kappa^{-1}}^2 = \int_D \kappa^{-1} \lvert\boldsymbol{i} \rvert^2 \, dv. \tag*{(19)}

Because the reference is the steady-ramp solution, eˉ\bar{e} initially includes the physical startup transient. However, according to Eq. (29) the continuum current discrepancy has fallen below 0.1% of its initial value by 25.2 μ\mus.

For both H(Curl)H(\mathrm{Curl}) formulations, the error settles within the first ten increments (50 μ\mus) to a level that decreases with each refinement. At t=0.15t = 0.15 ms, the order-zero errors are 16.07%, 8.31%, and 4.31%, from coarsest to finest. The corresponding order-one errors are 3.07%, 1.55%, and 0.895%. Both H(Curl)H(\mathrm{Curl}) formulations correctly predict that the eddy-current field becomes time-independent after the startup transient, even as the prescribed magnetic field continues to increase linearly, as in Eq. (17).

By contrast, the penalized H1H^1 current continues to evolve well beyond the physical startup transient and ultimately settles to an incorrect current distribution (Figure 3(b)). Mesh refinement initially reduces the current error but increases the duration of this erroneous transient. As shown in Figure 4, the error rises back above 30% at 0.040, 0.155, and 0.610 ms, respectively, from coarsest to finest. Thus, halving the element size delays this rise by approximately a factor of four at each refinement. At 1 ms, the errors on the two coarser meshes are nearly identical (43.60% and 43.65%), whereas the finest-mesh error is lower (36.95%) and still rising. Additional time increments confirm that the error on the finest mesh eventually reaches approximately the same plateau as on the coarser meshes. Refinement therefore delays the rise in error but does not correct the settled current distribution.

Discussion

Additional insight is obtained by comparing the present results with those for the TEAM 4 problem [18] and by considering the penalty-free order-one H1H^1 formulation. In the TEAM 4 problem, all element formulations considered in this work produce reasonable eddy-current solutions, including the penalty-free (λ=0\lambda= 0) H1H^1 formulation. By contrast, neither the penalized nor the penalty-free H1H^1 formulation reproduces the settled current distribution in the present benchmark. This suggests that the TEAM 4 problem does not strongly excite the spurious discrete gradient components identified in the preceding analysis, whereas the present benchmark is sensitive to them. These components belong to Πa(H(Curl))/Πa(ker⁡(∇×))∩∇dSϕ\Pi_a(H(\mathrm{Curl}))/\Pi_a(\ker(\nabla\times)) \cap\nabla_d S_\phi,13 which is nontrivial for the H1H^1 approximation considered here. The TEAM 4 results raise the possibility that many standard eddy-current benchmarks may not strongly excite this space. Agreement on such benchmarks alone could therefore give the impression that nodal elements generally entail only a minor loss of accuracy. In the present benchmark, however, the discrepancy persists under sustained loading and is apparent not only in the current magnitude but also in its direction at the exterior boundary (Figure 3). The pronounced normal component of the computed current is consistent with the loss of the discrete de Rham structure discussed previously: the continuum condition i⋅n=0\boldsymbol{i} \cdot\boldsymbol{n} = 0 follows from the prescribed tangential magnetic field in Eq. (15), without being imposed separately in the ϕ\phi equation.

Moreover, it is evident from Figure 5 that one cannot simply increase or remove the penalty parameter to improve the accuracy of the settled H1H^1 solution. On the 72-element mesh, removing the Coulomb gauge penalty (i.e., setting λ=0\lambda= 0) delays the rise back above 40% error from 0.080 to 0.180 ms. With λ=1000\lambda= 1000, which is 1000 times the conventional value, the error instead never drops below 39%. After the decay of the startup transient, the error rises almost immediately above 44% and then approaches the plateau of 43.6% from above, whereas the other two H1H^{1} solutions approach it from below. The current plots are likewise strikingly similar, with each H1H^{1} solution exhibiting a pronounced current component normal to the exterior faces. This example thus demonstrates that Coulomb gauge enforcement does not correct the underlying incompatibility of the approximation spaces, and that the failure to satisfy the implied condition i⋅n=0\boldsymbol{i} \cdot\boldsymbol{n} = 0 persists with or without the penalty term.

Figure 5. Effect of the Coulomb penalty on the order-one nodal H1H^{1} formulation on the 72-element mesh. (a) Relative current error of Eq. (18) at all 200 recorded increments. (b) The first 125 μ\mus of the same histories. (c,d) Top-face current-density magnitude and in-plane current vectors at 1 ms for the penalty-free and increased-penalty cases, respectively, with common magnitude and arrow scales.

Finally, it is worth noting that the order-one H(Curl⁡)H(\operatorname{Curl}) formulation on the coarsest mesh gives a smaller relative current error than the order-zero formulation on any of the meshes considered (Figure 4). At t=0.15t = 0.15 ms, the error on the 72-element mesh is 3.07% for the order-one formulation, compared with 16.07% for the order-zero formulation. Even on the 4608-element mesh, the order-zero error is 4.31%. For the same 30 time increments, the coarse order-one solution required only 4.14 CPU seconds, compared with 1237.40 CPU seconds for the finest order-zero solution. Thus, by the chosen error measure, the more accurate solution was obtained in approximately one three-hundredth of the computational time. This result is of particular interest because only the vector potential was enriched: the scalar potential retained its trilinear interpolation. Although the resulting mixed H1/H(Curl⁡)H^1/H(\operatorname{Curl}) pair is not exact, it preserves the lower-order discrete subcomplex. The present example thus emphasizes that, in time-domain problems of this type, preserving this subcomplex may be more important for current accuracy than requiring the mixed pair itself to be exact, potentially allowing for more efficient and accurate computation of solutions when one field requires a richer approximation than the other.

Conclusion

In this work, the time-domain aa-ϕ\phi electromagnetics equations and associated finite element implementations have been concisely summarized, and the issues plaguing H1H^1 continuous vector interpolations of the magnetic vector potential have been stated in their simplest form. This explanation is useful given that many similar explanations can be esoteric and difficult to quickly parse. Of primary importance is the clear demonstration of the failure of H1H^1 continuous vector-potential interpolations with penalty regularization of the (Coulomb) gauge in a simple benchmark problem. Although the H1H^1 solution initially appears to approach the semi-analytic reference, it ultimately settles to an incorrect current distribution that violates the no-through-current condition. Mesh refinement delays the approach to this state but does not correct it, nor does increasing or removing the penalty. In contrast, both H(Curl⁡)H(\operatorname{Curl}) formulations predict the expected time-independent eddy-current response within the first ten increments, and their settled current errors decrease with refinement. Moreover, the order-one formulation on the coarsest mesh gives a smaller relative current error than the order-zero formulation on the finest mesh in approximately one three-hundredth of the CPU time. Although this example demonstrates that exactness of the full mixed pair is not strictly necessary for an accurate current, both H(Curl⁡)H(\operatorname{Curl}) formulations preserve an exact lower-order discrete subcomplex, whereas the nodal H1H^1 formulation lacks this structure and fails to recover the correct settled current even under mesh refinement.

Construction of the reference solution

Coordinates in the series and edge integrals below are given in millimeters, with all other units in SI.

Analytic solution form

To solve Eq. (16), extend the four sides of the Ti core to the perimeter to create nine square subdomains of equal size. By symmetry, only the three squares outlined in Figure 6 need to be explicitly considered. Denote the lower-left corner of the square being evaluated by (xc,yc)(x_c,y_c). Its coordinates are (−0.75,−0.75)(-0.75,-0.75), (−0.25,−0.75)(-0.25,-0.75), and (−0.25,−0.25)(-0.25,-0.25) for the corner, side, and core, respectively, with x=y=0x=y=0 at the center of the full cross section.

Partition of the cross section showing the corner, side, and core subdomains

Figure 6. Partition of the cross section superimposed on the finite element mesh.

On each square, write g=pκ+hg=p_\kappa+h, where pκp_\kappa solves −div⁡∇pκ=μ0βκ-\operatorname{div}\nabla p_\kappa=\mu_0\beta\kappa with pκ=0p_\kappa=0 on all four sides. The harmonic correction satisfies div⁡∇h=0\operatorname{div}\nabla h=0 with h=gh=g on the sides of the square. The particular solution is the same on all three squares up to a translation and the value of κ\kappa, and is given below following Haberman [14]:

pκ(x,y)=10−6μ0βκ∑n≥1n oddsin⁡(2nπ[x−xc])(nπ)3[1−cosh⁡(2nπ[y−yc−0.25])cosh⁡(nπ/2)].(20)p_\kappa(x,y)=10^{-6}\mu_0\beta\kappa\sum_{\substack{n\ge1\\ n\ \mathrm{odd}}}\frac{\sin(2n\pi[x-x_c])}{(n\pi)^3}\left[1-\frac{\cosh(2n\pi[y-y_c-0.25])}{\cosh(n\pi/2)}\right]. \tag*{(20)}

In addition, hh can be written as the bilinear interpolant B(x,y)B(x,y) of the corner values of gg plus a weighted sum of edge modes. By symmetry, B=g=gcB=g=g_c at each corner of the core, and from the applied boundary conditions B=g=0B=g=0 at each corner on the outer perimeter.

The four edge modes for each nn are rotations of the top-edge mode up to orientation:

Hntop(x,y)=sin⁡(2nπ[x−xc])sinh⁡(2nπ[y−yc])sinh⁡(nπ),H^{\mathrm{top}}_n(x,y)=\sin\left(2n\pi[x-x_c]\right)\frac{\sinh\left(2n\pi[y-y_c]\right)}{\sinh(n\pi)},
Hnbottom(x,y)=sin⁡(2nπ[x−xc])sinh⁡(2nπ[yc+0.5−y])sinh⁡(nπ),H^{\mathrm{bottom}}_n(x,y)=\sin\left(2n\pi[x-x_c]\right)\frac{\sinh\left(2n\pi[y_c+0.5-y]\right)}{\sinh(n\pi)},
Hnright(x,y)=sin⁡(2nπ[y−yc])sinh⁡(2nπ[x−xc])sinh⁡(nπ),H^{\mathrm{right}}_n(x,y)=\sin\left(2n\pi[y-y_c]\right)\frac{\sinh\left(2n\pi[x-x_c]\right)}{\sinh(n\pi)},
Hnleft(x,y)=sin⁡(2nπ[y−yc])sinh⁡(2nπ[xc+0.5−x])sinh⁡(nπ),(21)H^{\mathrm{left}}_n(x,y)=\sin\left(2n\pi[y-y_c]\right)\frac{\sinh\left(2n\pi[x_c+0.5-x]\right)}{\sinh(n\pi)}, \tag*{(21)}

where the edge coefficients are zero on the outer perimeter. Let ana_n denote the coefficients on the Al–Al cuts and bnb_n those on the Al–Ti interfaces. On each Al–Ti interface, g−gcg-g_c is symmetric about the midpoint, whereas the even-indexed sine modes are antisymmetric. Thus, bn=0b_n=0 for even nn, and the analytic solution on each square subset is

gcorner(x,y)=pκa+4gc(x−xc)(y−yc)+lim⁡N→∞∑n=1Nan(Hntop+Hnright),(22)g_{\mathrm{corner}}(x,y)=p_{\kappa_a}+4g_c(x-x_c)(y-y_c)+\lim_{N\to\infty}\sum_{n=1}^{N}a_n\left(H^{\mathrm{top}}_n+H^{\mathrm{right}}_n\right), \tag*{(22)}
gside(x,y)=pκa+2gc(y−yc)+lim⁡N→∞∑n=1N[an(Hnleft+Hnright)+bnHntop],(23)g_{\mathrm{side}}(x,y)=p_{\kappa_a}+2g_c(y-y_c)+\lim_{N\to\infty}\sum_{n=1}^{N}\left[a_n\left(H^{\mathrm{left}}_n+H^{\mathrm{right}}_n\right)+b_nH^{\mathrm{top}}_n\right], \tag*{(23)}
gcore(x,y)=pκa10+gc+lim⁡N→∞∑n=1Nbn(Hnbottom+Hntop+Hnleft+Hnright).(24)g_{\mathrm{core}}(x,y)=\frac{p_{\kappa_a}}{10}+g_c+\lim_{N\to\infty}\sum_{n=1}^{N}b_n\left(H^{\mathrm{bottom}}_n+H^{\mathrm{top}}_n+H^{\mathrm{left}}_n+H^{\mathrm{right}}_n\right). \tag*{(24)}

Here, the remaining unknowns gcg_c, ana_n, and the odd-indexed bnb_n can be determined from the required tangential continuity of ee, which manifests as normal continuity of κ−1∇g\kappa^{-1}\nabla g.

Determination of the coefficients

The tile solutions in Eqs. (22)–(24) agree in value on each shared edge by construction, so it remains to match κ−1∇g⋅n\kappa^{-1}\nabla g\cdot\mathbf{n}. By symmetry, it suffices to consider one of the eight Al–Al cuts and one of the four Al–Ti interfaces. For the squares outlined in Figure 6, these are chosen to lie on x=−0.25x=-0.25 and y=−0.25y=-0.25, respectively. Let ww denote a scalar H1H^1 function defined on the full cross section that vanishes on the outer perimeter and has the same symmetries as gg. For every such ww, the interface condition can be written in weak form as

8∫−0.75−0.25(∂gcorner∂x−∂gside∂x)x=−0.25w(−0.25,y) dy+4∫−0.250.25(∂gside∂y−10∂gcore∂y)y=−0.25w(x,−0.25) dx=0.(25)8\int_{-0.75}^{-0.25}\left(\frac{\partial g_{\mathrm{corner}}}{\partial x}-\frac{\partial g_{\mathrm{side}}}{\partial x}\right)_{x=-0.25}w(-0.25,y)\,\mathrm{d}y+4\int_{-0.25}^{0.25}\left(\frac{\partial g_{\mathrm{side}}}{\partial y}-10\frac{\partial g_{\mathrm{core}}}{\partial y}\right)_{y=-0.25}w(x,-0.25)\,\mathrm{d}x=0. \tag*{(25)}

Here, the factor of ten accounts for the lower conductivity of the core.

In order to compute the reference solution, the edge series is truncated at n=Nn=N. The test space is chosen in a Bubnov–Galerkin sense: it is spanned by the functions multiplying gcg_c, ana_n, and bnb_n in the edge values of gg. Written as pairs (w(−0.25,y),w(x,−0.25))(w(-0.25,y), w(x,-0.25)), these are

(2(y+0.75),1),(sin⁡(2nπ[y+0.75]),0),(0,sin⁡(2nπ[x+0.25])),(2(y+0.75),1), \qquad(\sin(2n\pi[y+0.75]),0), \qquad(0,\sin(2n\pi[x+0.25])),

respectively, with 1≤n≤N1 \le n \le N for ana_n and only odd n≤Nn \le N for bnb_n. Evaluating the integrals in Eq. (25) analytically for each test function gives one linear equation per coefficient, with the known terms from pκp_\kappa supplying the right-hand side. The resulting symmetry-reduced system is then solved by LU factorization using numpy.linalg.solve.

To compute the reference solution used in § 5–§ 6, the coefficient system is solved with N=512N = 512, while pκp_\kappa is evaluated separately according to Eq. (20) by summing the purely sinusoidal part exactly and retaining the odd modes through n=2047n = 2047 in the hyperbolic part. With this cutoff to pκp_\kappa held fixed, doubling NN from 256 to 512 changes the current by only 0.0685% relative to the 512-mode solution in the norm of Eq. (19). The reference current is therefore taken to be converged for the comparisons presented in this work.

Comparison with the ramp started from zero

The simulations start from zero current, whereas the steady-ramp current in Eq. (17) is independent of time. Letting rr denote the difference between the sudden-ramp solution associated with the simulations and the reference solution presented above, the linearity of the differential operators and identical tangential b\boldsymbol{b}-field boundary datum imply that

μ0∂tr=div⁡(κ−1∇r),r=0 on the perimeter,r(x,y,0)=−g(x,y).(26)\mu_0 \partial_t r = \operatorname{div}\left(\kappa^{-1}\nabla r\right), \qquad r = 0 \text{ on the perimeter}, \qquad r(x,y,0)=-g(x,y). \tag*{(26)}

Applying the relative error of Eq. (18) to the continuum current and using Eq. (17) gives

eˉ2(t)=∫Dκ−1∣∇r∣2 dv∫Dκ−1∣∇g∣2 dv=∫κ−1∣∇r∣2 da∫κ−1∣∇g∣2 da.\bar e^2(t)=\frac{\int_{\mathcal{D}}\kappa^{-1}|\nabla r|^2\,dv}{\int_{\mathcal{D}}\kappa^{-1}|\nabla g|^2\,dv}=\frac{\int\kappa^{-1}|\nabla r|^2\,da}{\int\kappa^{-1}|\nabla g|^2\,da}.

The second equality follows because rr, gg, and κ\kappa are independent of zz, so the thickness contributions cancel. Since gg is independent of time, differentiation gives

∂∂teˉ2(t)=2eˉ(t)[eˉ(t)],t=2∫κ−1∇r⋅∇r,t da∫κ−1∣∇g∣2 da.\frac{\partial}{\partial t}\bar e^2(t)=2\bar e(t)[\bar e(t)]_{,t}=\frac{2\int\kappa^{-1}\nabla r\cdot\nabla r_{,t}\,da}{\int\kappa^{-1}|\nabla g|^2\,da}.

Dividing both sides by −2eˉ2(t)-2\bar e^2(t), applying the 2D-divergence theorem, and noting Eq. (26) then yields

−[eˉ(t)],teˉ(t)=−∫κ−1∇r⋅∇r,t da∫κ−1∣∇r∣2 da=∫[r,tdiv⁡(κ−1∇r)−div⁡(r,tκ−1∇r)] da∫κ−1∣∇r∣2 da=μ0∫(r,t)2 da∫κ−1∣∇r∣2 da.(27)-\frac{[\bar e(t)]_{,t}}{\bar e(t)} =-\frac{\int\kappa^{-1}\nabla r\cdot\nabla r_{,t}\,da}{\int\kappa^{-1}|\nabla r|^2\,da} =\frac{\int\left[r_{,t}\operatorname{div}\left(\kappa^{-1}\nabla r\right)-\operatorname{div}\left(r_{,t}\kappa^{-1}\nabla r\right)\right]\,da}{\int\kappa^{-1}|\nabla r|^2\,da} =\frac{\mu_0\int(r_{,t})^2\,da}{\int\kappa^{-1}|\nabla r|^2\,da}. \tag*{(27)}

To establish a bound on eˉ(t)\bar e(t) from above, it suffices to bound −[eˉ(t)],t/eˉ(t)-[\bar e(t)]_{,t}/\bar e(t) in Eq. (27) from below. First, note that Eq. (26) implies

μ0∫rr,t da=∫rdiv⁡(κ−1∇r) da=−∫κ−1∣∇r∣2 da.\mu_0\int r r_{,t}\,da=\int r\operatorname{div}\left(\kappa^{-1}\nabla r\right)\,da=-\int\kappa^{-1}|\nabla r|^2\,da.

By Cauchy–Schwarz and the identity just obtained,

(∫κ−1∣∇r∣2 da)2=μ02(∫rr,t da)2≤μ02(∫r2 da)(∫(r,t)2 da).\left(\int\kappa^{-1}|\nabla r|^2\,da\right)^2=\mu_0^2\left(\int r r_{,t}\,da\right)^2\leq\mu_0^2\left(\int r^2\,da\right)\left(\int(r_{,t})^2\,da\right).

Dividing both sides by μ0(∫r2 da)(∫κ−1∣∇r∣2 da)\mu_0\left(\int r^2\,da\right)\left(\int\kappa^{-1}|\nabla r|^2\,da\right) gives

∫κ−1∣∇r∣2 daμ0∫r2 da≤μ0∫(r,t)2 da∫κ−1∣∇r∣2 da=−[e‾(t)],te‾(t),\frac{\int\kappa^{-1}|\nabla r|^2\,da}{\mu_0\int r^2\,da}\leq\frac{\mu_0\int(r_{,t})^2\,da}{\int\kappa^{-1}|\nabla r|^2\,da}=-\frac{[\overline{e}(t)]_{,t}}{\overline{e}(t)},

where the equality follows from Eq. (27).

The boundary data allow us to bound the quotient on the left by a known constant. Indeed, since rr vanishes at both ends of each horizontal line through the square, the one-dimensional Poincaré inequality gives

∫r2 da≤(1.5×10−3)2π2∫(r,x)2 da.\int r^2\,da\leq\frac{(1.5\times10^{-3})^2}{\pi^2}\int(r_{,x})^2\,da.

Applying the same inequality along the vertical lines gives an analogous bound involving r,yr_{,y}. Adding these two bounds, dividing by two, and using κ≤κa\kappa\leq\kappa_a gives

∫r2 da≤(1.5)2×10−62π2∫∣∇r∣2 da≤1.125×10−6κaπ2∫κ−1∣∇r∣2 da.\int r^2\,da\leq\frac{(1.5)^2\times10^{-6}}{2\pi^2}\int|\nabla r|^2\,da\leq\frac{1.125\times10^{-6}\kappa_a}{\pi^2}\int\kappa^{-1}|\nabla r|^2\,da.

Rearranging yields the desired bound:

106π21.125μ0κa≤∫κ−1∣∇r∣2 daμ0∫r2 da≤−[e‾(t)],te‾(t).(28)\frac{10^6\pi^2}{1.125\mu_0\kappa_a}\leq\frac{\int\kappa^{-1}|\nabla r|^2\,da}{\mu_0\int r^2\,da}\leq-\frac{[\overline{e}(t)]_{,t}}{\overline{e}(t)}. \tag*{(28)}

The current starts from zero, so e‾(0)=1\overline{e}(0)=1. Integrating the right-hand side of Eq. (28) from zero to tt therefore gives −log⁡e‾(t)-\log\overline{e}(t), and hence

e‾(t)≤exp⁡[−106π21.125μ0κat].(29)\overline{e}(t)\leq\exp\left[-\frac{10^6\pi^2}{1.125\mu_0\kappa_a}t\right]. \tag*{(29)}

The time constant in this bound is 3.6354 μ\mus. Thus, by 25.2 μ\mus the continuum current is within 0.1% of the steady-ramp current in the norm of Eq. (19).

References

  1. [2]Arnold, D.N., Awanou, G., 2014. Finite element differential forms on cubical meshes. Mathematics of Computation 83, 1551–5170.DOI
  2. [3]Balay, S., Abhyankar, S., Adams, M.F., Benson, S., Brown, J., Brune, P., Buschelman, K., Constantinescu, E., Dalcin, L., Dener, A., Eijkhout, V., Faibussowitsch, J., Gropp, W.D., Hapla, V., Isaac, T., Jolivet, P., Karpeev, D., Kaushik, D., Knepley, M.G., Kong, F., Kruger, S., May, D.A., McInnes, L.C., Mills, R.T., Mitchell, L., Munson, T., Roman, J.E., Rupp, K., Sanan, P., Sarich, J., Smith, B.F., Suh, H., Zampini, S., Zhang, H., Zhang, H., Zhang, J., 2024. PETSc/TAO Users Manual. Technical Report ANL-21/39 - Revision 3.22. Argonne National Laboratory.
  3. [4]Balay, S., Gropp, W.D., McInnes, L.C., Smith, B.F., 1997. Efficient management of parallelism in object oriented numerical software libraries, in: Arge, E., Bruaset, A.M., Langtangen, H.P. (Eds.), Modern Software Tools in Scientific Computing, Birkhäuser Press. pp. 163–202.DOI
  4. [5]Boffi, D., Fernandes, P., Gastaldi, L., Perugia, I., 1999. Computational models of electromagnetic resonators: Analysis of edge element approximation. SIAM Journal on Numerical Analysis 36, 1264–1290.
  5. [6]Bossavit, A., 1998a. On the geometry of electromagnetism. (1): Euclidean space. Journal of the Japan Society of Applied Electromagnetics and Mechanics 6, 17–28.
  6. [10]Carlson, T., Govindjee, S., 2026. A fully coupled, electrochemical-magnetic theory with applications to solid-state battery materials. Journal of the Mechanics and Physics of Solids 209, 106514.
  7. [11]Cendes, Z., 1991. Vector finite elements for electromagnetic field computation. IEEE Transactions on Magnetics 27, 3958–3966.DOI
  8. [14]Haberman, R., 2004. Applied Partial Differential Equations: With Fourier Series and Boundary Value Problems. 4th ed., Pearson Prentice Hall.
  9. [15]Hiptmair, R., 2002. Finite elements in computational electromagnetism. Acta Numerica 11, 237–339.
  10. [16]Jackson, J.D., 2021. Classical Electrodynamics, International Adaptation. 3rd ed., John Wiley & Sons, Incorporated, New York.
  11. [18]Kameari, A., 1988. Results for benchmark calculations of problem 4 (the felix brick). COMPEL 7, 65–80.DOI
  12. [19]Kamireddy, D., Chavan, S.M., Nandy, A., 2022. Comparative performance of novel nodal-to-edge finite elements over conventional nodal element for electromagnetic analysis. URL: https://arxiv.org/abs/2203.14522, arXiv:2203.14522.
  13. [20]Kovetz, A., 2000. Electromagnetic Theory. Oxford University Press, New York, NY.
  14. [21]Kuczmann, M., 2008. Nodal and edge finite element analysis of eddy current field problems. Przegląd Elektrotechniczny 84, 194–197.
  15. [22]Lee, J.M., 2012. Introduction to Smooth Manifolds. Graduate Texts in Mathematics. 2nd ed., Springer Science+Business Media New York, NY.
  16. [25]Nakata, T., Takahashi, N., Fujiwara, K., Shiraki, Y., 1990. Comparison of different finite elements for 3-d eddy current analysis. IEEE Transactions on Magnetics 26, 434–437.DOI
  17. [26]Stewart, E.M., Anand, L., 2026. Electromagneto-elasticity in the eddy current approximation and its numerical implementation in the open-source finite element program FEniCSx. Computational Mechanics (early view) https://doi.org/10.1007/s00466-026-02791-w.
  18. [27]Taylor, R.L., Govindjee, S., 2026. FEAP – A Finite Element Analysis Program. URL: http://feap.berkeley.edu/wiki/index.php/Manuals.
  19. [28]Zaglmayr, S., 2006. High Order Finite Element Methods for Electromagnetic Field Computation. Ph.D. thesis. Johannes Kepler University.

Paper details

vlnc:W_wh8qw24f
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 7

downloads · 0

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

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents