betaFeedback

WorksComputer Science

Certificate-Based Verification of Pulsed Field Ablation Models

Abstract

Computational models of pulsed field ablation predict the lesion as the tissue where the electric field exceeds a lethal threshold. Their numerical error is rarely bounded, and automated pipelines increasingly run them without an expert checking each result. To verify each run, we separate the question "was the stated model solved correctly?" from the solver that answers it. An untrusted solver returns with its answer a certificate, consisting of a potential with the correct electrode and ground values and an exactly charge-conserving current. A small independent checker verifies the certificate against a written specification of the physics. Using the Prager-Synge hypercircle theorem evaluated in ball arithmetic, it turns the certificate into guaranteed two-sided bounds on the impedance and on the root-mean-square field of every tile of a lesion map, also over a range of blood/tissue conductivity ratios. A vectorized version of the checker takes about a tenth of the solve time. For a focal electrode in contact, the tile-resolved depth at 1000 V and 700 V/cm is proven to lie in 3.50-3.75 mm for every ratio from 1.7 to 3.75. For a bipolar catheter, certified depths overlap published in-vivo depths. Where they differ from a published reference model, the certificates prove that the difference is not numerical error in our computation. We compare the certificate with a grid convergence study on the same solver. On graded grids the study is sharper for local fields but remains an estimate. On uniform grids it fails to converge at the electrode edge, where the certificate still gives a valid bracket. Solvers that deliberately solve equations other than the specified ones pass the grid convergence study, but the certificates refute their claims or leave them unsupported.

Introduction

Pulsed field ablation (PFA) kills myocardium by irreversible electroporation where the electric field exceeds a lethal threshold, so lesion size is predicted by solving a quasi-static conduction problem and taking a threshold contour [1, 2]. Finite-element simulation is the standard tool for mapping the field of a PFA catheter, while the lethal threshold and tissue response remain the main sources of model uncertainty [3]. Such models are used to compare catheter geometries, to choose voltages, and, increasingly, to support regulatory submissions. The verification and validation (V&V) standard V&V 40 of the American Society of Mechanical Engineers and the guidance of the U.S. Food and Drug Administration on computational modeling set risk-informed credibility goals: the more a decision relies on a model, and the graver its consequence, the more evidence is needed [4, 5].

On the numerical side, evidence for accuracy comes at three levels [6, 7]. Code verification, for example by manufactured solutions and order-of-accuracy tests, shows that the software can solve its equations correctly. Calculation verification estimates the discretization, solver, and round-off error of the simulations actually used. It is commonly a mesh refinement study (mesh study for short) on a representative case, with the result assumed to carry over to the runs in the context of use. It gives estimates rather than bounds, and repeating it for every run of a design loop is costly. Guaranteed a posteriori bounds make calculation verification rigorous. They prove a bound on the error of a particular run, a small program independent of the solver can check them, and they can be attached to every run rather than to a representative one. This paper concerns that level.

Estimates have been acceptable because an expert chose the mesh and checked each result. Simulation is increasingly automated, by scripted design loops, by solvers written with artificial intelligence, and by learned surrogates, i.e., machine-learning models such as neural operators that predict full fields in seconds [8]. Automated simulations are not known to be less accurate, but once the expert leaves the loop, the per-run check goes with them. This paper does not test solvers written by artificial intelligence or learned surrogates. Solvers with deliberately wrong physics stand in for the former, and surrogates are discussed in § 7.

We take the view of proof-carrying code [9] and of certifying algorithms [10]. In this view, the solver is untrusted, and its output carries evidence that a small checker can verify (§ 3). For linear conduction the evidence is classical. The hypercircle of Prager and Synge [11, 12] bounds the energy error of any admissible potential by its distance to any admissible, exactly conserved current. The complementary Dirichlet and Thomson principles give two-sided bounds on capacities and impedances [13]. Equilibrated-flux estimators built on these ideas are mature [14, 15, 16]. Functional majorants extend them to arbitrary approximations and to problems with uncertain data [17, 18]. Verified computation combines hypercircle bounds with interval arithmetic to prove statements about partial differential equations (PDEs) [19, 20, 21]. Certified reduced-basis methods bound outputs over parameter domains [22], and neural PDE solvers have been given error majorants [23]. Equilibrated estimators have been extended to parametric problems [24] and to include the algebraic error of inexact solves [25], and they certify existence and outputs for semilinear problems by post-processing a conventional finite-element solution, there evaluated in floating point [26]. Checking the output of an untrusted solver by an independent program is established in exact optimization [27].

None of these ingredients is new. Our contribution is their assembly into a small checker, independent of the solver, that turns any admissible potential-current pair into guaranteed bounds, rounding included, on the impedance and the tile-resolved lesion map of a PFA electrode, also over ranges of conductivity. Solvers whose equations deliberately differ from the specification converge under refinement and pass a grid convergence study, but the checker refutes their claims or leaves them unsupported. To the best of our knowledge, these are the first lesion maps for PFA with proven numerical error bounds. In particular, the paper provides:

  • A written specification of the physics, kept separate from every numerical choice, against which certificates are checked.

  • Exact integration in ball arithmetic [28], so the bounds are proven rather than estimated, including rounding.

  • A vectorized checker in double-precision ball arithmetic, rigorous under stated assumptions, that checks the lesion map in about a tenth of the solve time.

  • Certified lesion maps (proven lethal, proven spared, undecided), stated for the root-mean-square field of each tile rather than pointwise (§ 2).

  • Bounds that hold over a range of tissue conductivities, obtained by blending a few certificates.

  • A comparison with a standard grid convergence study on the same solver, including solvers with deliberately wrong physics that the study accepts and the certificates do not.

  • A first independence check: the unchanged checker certifies a potential computed by FeniCSx on the same grid, with bounds identical to those for our solver (Table 7).

Solver of the potentialCertified impedance (Ω\Omega)η\etaTile width, median / worst
Our solver93.950274296–93.9502757891.3×10−51.3 \times10^{-5}1.9% / 18%
FEniCSx 0.1193.950274296–93.9502757891.3×10−51.3 \times10^{-5}1.9% / 18%

Table 7. Monopolar lesion map, contact, nominal ratio; same grid and degree-3 tensor elements; current from the stream-function fit in both rows. Bounds rounded outward.

Certificates bound the numerical error of the stated model. Whether the model and threshold describe tissue is a separate question of validation, discussed in § 7.

Model and Lesion Criterion

Geometry and Equations

We use cylindrical coordinates (r,z)(r,z) with rotational symmetry and the measure dμ=2πr dr dz\mathrm{d}\mu= 2\pi r\,\mathrm{d}r\,\mathrm{d}z. Myocardium (conductivity σt\sigma_t) fills z<0z < 0 and blood (σb\sigma_b) fills z>0z > 0. A flat-tipped cylindrical metal electrode of radius a=1.25a = 1.25 mm and length LL sits a gap g≥0g \ge0 above the tissue and continues as an insulating shaft. In the monopolar configuration the electrode is held at VV and the outer boundary, 50 mm away, is grounded. In the bipolar configuration an insulating spacer separates a tip electrode from a ring electrode. The generator floats, so the outer boundary is insulating and only the potential difference VV matters; we hold the tip at +V/2+V/2 and the ring at −V/2-V/2 without loss of generality. Figure 1 shows both configurations. The potential uu solves

Axisymmetric monopolar and bipolar electrode models

Figure 1. Axisymmetric model in the (r,z)(r,z) half-plane, not to scale. (a) Monopolar: the electrode is held at VV and the outer boundary, 50 mm away, is grounded. (b) Bipolar: tip and ring electrodes at ±V/2\pm V/2 are separated by an insulating spacer, and the outer boundary is insulating. The axis is a symmetry line, and the electrode sits a gap g≥0g \ge0 above the tissue.

−∇⋅(σ∇u)=0in the domain,-\nabla\cdot(\sigma\nabla u) = 0 \quad\text{in the domain,}
u=uDon ΓD,u = u_D \quad\text{on } \Gamma_D,
σ∂nu=0on ΓN,(1)\sigma\partial_n u = 0 \quad\text{on } \Gamma_N, \tag*{(1)}

with ΓD\Gamma_D the electrodes (and ground) and ΓN\Gamma_N the insulating walls and the axis. All boundaries are axis-aligned, so the problem lives on a rectangular grid. Electrodes are voltage-controlled and conductivities are constant, so the problem is linear in VV and one certificate at 1 V serves every voltage.

Lesion Criterion

A lesion-map tile ω\omega is a ring of square cross-section, 0.25 mm wide, in one material. It is lethal when its root-mean-square (RMS) field Eω=(∫ω∣∇u∣2 dμ/∣ω∣)1/2E_\omega= \left(\int_\omega|\nabla u|^2\,\mathrm{d}\mu/|\omega|\right)^{1/2} reaches the threshold EthE_{\mathrm{th}}. Given a certified interval for EωE_\omega, a tile is proven lethal, proven spared, or undecided. We report two depths. The depth “reached” is the bottom of the deepest axis tile such that it and every axis tile above it are proven lethal; it bounds the deepest lethal tile from below. The depth “not beyond” is the level below which every tile of the map, in every column, is proven spared; the statement is limited to the mapped region. Lesion radii are defined in the same way along each row of tiles. The axis tiles are the smallest and carry the widest intervals. The proven-lethal inner envelope and the proven-spared outer envelope bound the tile-resolved lesion from inside and outside.

This criterion coarsens the pointwise threshold contour used by Meckes et al. and Petras et al. [1, 2]. In a proven-lethal tile ∣∇u∣≥Eth\lvert\nabla u\rvert\ge E_{\mathrm{th}} on a set of positive measure, and in a proven-spared tile ∣∇u∣<Eth\lvert\nabla u\rvert< E_{\mathrm{th}} on a set of positive measure. Neither statement holds for every point of the tile, so a proven-spared tile may still contain points above the threshold, and a safety claim about collateral tissue needs pointwise bounds (§ 7). The envelopes locate the pointwise contour to within about one tile where the field decreases monotonically along the axis. We use 700 V/cm as an illustrative threshold within the range 270–1000 V/cm that Petras et al. find for depth and width [2], and 268 V/cm where we follow Meckes et al.

Certificates and the Checker

Figure 2 shows how a solver’s output becomes proven bounds.

From solver to proven bounds

Figure 2. From solver to proven bounds. Orange: untrusted; blue: trusted. The solver returns a certificate, an admissible potential vv and a stream function ψ\psi whose curl is an exactly charge-conserving current. The checker never runs the solver. A certificate that fails the specification check is rejected. Otherwise numerical verification returns bounds that hold for the stated model whatever the solver did, and claims are judged against them. Tiles are judged by their RMS field, not pointwise. A solver of the wrong equations passes the first test, but the bounds refute its claims or leave them unsupported. A mesh study compares a solver only with itself and makes neither test. Certificates at a few conductivity ratios are blended into bounds that hold for every ratio in between (§ 4).

The Hypercircle

Write ∥X∥2=∫∣X∣2σ−1 dμ\lVert X\rVert^2 = \int\lvert X\rvert^2\sigma^{-1}\,\mathrm{d}\mu, with ∥⋅∥ω\lVert\cdot\rVert_\omega the same integral restricted to ω\omega. A potential v∈H1v \in H^1 (with the weight rr) is admissible if v=uDv = u_D on ΓD\Gamma_D. A current F∈H(div⁡)F \in H(\operatorname{div}) is admissible if ∇⋅F=0\nabla\cdot F = 0 and F⋅n=0F\cdot n = 0 on ΓN\Gamma_N.

Theorem 1 (Prager–Synge [11]). For admissible vv and FF, ∥σ∇v−σ∇u∥2+∥σ∇u−F∥2=∥σ∇v−F∥2=:η2\|\sigma\nabla v-\sigma\nabla u\|^2+\|\sigma\nabla u-F\|^2=\|\sigma\nabla v-F\|^2=:\eta^2.

Proof. Let w=v−uw=v-u, which vanishes on ΓD\Gamma_D. Then ∫∇w⋅(σ∇u−F) dμ=0\int\nabla w\cdot(\sigma\nabla u-F)\,\mathrm{d}\mu=0. The first term vanishes by the weak form of the problem. In the second, integration by parts leaves a volume term, which vanishes because ∇⋅F=0\nabla\cdot F=0, and boundary terms, which vanish because w=0w=0 on ΓD\Gamma_D and F⋅n=0F\cdot n=0 on ΓN\Gamma_N. The identity follows from Pythagoras’ theorem. □\square

Hence, σ∇u\sigma\nabla u lies on the sphere with center c=(σ∇v+F)/2c=(\sigma\nabla v+F)/2 and radius η/2\eta/2. Restricting to a tile ω\omega in one material and using the triangle inequality gives ∥σ∇u∥ω∈(∥c∥ω±η/2)\|\sigma\nabla u\|_\omega\in\bigl(\|c\|_\omega\pm\eta/2\bigr) (the lower end clipped at zero), so

Eω∈[(∥c∥ω−η/2)+, ∥c∥ω+η/2]/σω∣ω∣.(2)E_\omega\in\left[\left(\|c\|_\omega-\eta/2\right)_+,\ \|c\|_\omega+\eta/2\right]/\sqrt{\sigma_\omega|\omega|}. \tag*{(2)}

The radius here is the global η\eta: the bound is not localized, and its half-width η/(2σω∣ω∣)\eta/(2\sqrt{\sigma_\omega|\omega|}) is largest for the smallest tiles, those on the axis.

For the conductance G=∥σ∇u∥2/V2G=\|\sigma\nabla u\|^2/V^2, the Dirichlet principle gives G≤∥σ∇v∥2/V2G\le\|\sigma\nabla v\|^2/V^2, and the sphere gives (∥c∥−η/2)2/V2≤G≤(∥c∥+η/2)2/V2(\|c\|-\eta/2)^2/V^2\le G\le(\|c\|+\eta/2)^2/V^2. Let Φ\Phi be the current that FF carries off the driven electrode. In the bipolar case this is the tip, and the current through the ring is −Φ-\Phi. In both configurations ⟨F,σ∇u⟩=VΦ\langle F,\sigma\nabla u\rangle=V\Phi, so Cauchy–Schwarz gives Thomson’s bound G≥Φ2/∥F∥2G\ge\Phi^2/\|F\|^2. For the bipolar catheter, GG is the tip-to-ring conductance.

Certificate Format

On each grid cell, vv is a tensor polynomial of degree pp in (r,z)(r,z). The current is FF, the curl of a Stokes stream function ψ\psi, which is a tensor polynomial of degree pp in (r2,z)(r^2,z):

F=(−1r∂zψ,1r∂rψ).(3)F=\left(-\frac{1}{r}\partial_z\psi,\frac{1}{r}\partial_r\psi\right). \tag*{(3)}

This FF is divergence-free by construction, and continuity of ψ\psi gives continuity of the normal current across cells. The flux of FF through a wall is 2π2\pi times the change of ψ\psi along it. So F⋅n=0F\cdot n=0 on ΓN\Gamma_N exactly when ψ\psi is constant along each connected insulating wall. The axis belongs to such a wall with ψ=0\psi=0. For the floating bipolar catheter, the axis, the outer boundary, and the shaft above the ring form a single wall with ψ=0\psi=0. The spacer between the electrodes keeps its own constant, which is Φ/(2π)\Phi/(2\pi). Regularity on the axis needs both ψ=0\psi=0 there and the variable r2r^2; with rr in its place, FzF_z would behave like 1/r1/r. The certificate consists of nodal values on each cell, equispaced in rr for vv and in r2r^2 for ψ\psi. The solver may compute them in any way.

The Checker

The checker reads the specification and the nodal arrays, and never runs a solver. It makes two tests. The specification check consists of steps 1 and 2 below, and numerical verification of steps 3 and 4:

  1. It verifies the grid against the geometry.

  1. It verifies admissibility exactly: the electrode, ring, and ground values of vv, and ψ\psi zero or constant along each insulating wall, at every boundary node, including those inside cell edges (a degree-pp polynomial constant at the p+1p+1 nodes of an edge is constant on that edge).

  1. It rebuilds each cell’s polynomials by an exact change of basis.

  1. It integrates every inner product exactly.

Each integrand is a polynomial plus a term N(ξ)/(r1+hrξ)N(\xi)/(r_1+h_{r}\xi) from the 1/r1/r in FF. Near the axis this term is handled by exact polynomial division. On thin cells, where division would lose precision to cancellation, it is handled by moments obtained from a downward recurrence with a rigorous starting enclosure. All arithmetic is in 256-bit balls [28], and every float in the certificate is read as the exact binary number it is. The trusted base is the specification (about 220 lines of Python), the checker and range checker (about 750 lines), the ball-arithmetic library python-flint, and NumPy, which is used for the boundary masks and the exact equality tests. The code also implements primal-dual output bounds [29] for axis voxel means, which are not used in the results below.

Outcomes. The specification check decides whether a certificate is admissible. A run is rejected if it is not, for example when a boundary value differs from the specified one, and then no bounds follow. For an admissible certificate, numerical verification always returns proven bounds for the stated model, which include the discretization, solver, and rounding errors of the run. Claims are then judged against these bounds. A solver may report its own value of a certified quantity, here the impedance; we call this value its claim. The claim is refuted if it lies outside the certified interval, and unsupported if the interval is too wide to prove it to the required accuracy. The last outcome occurs for an inaccurate solver and for a solver of the wrong equations alike. Otherwise the run is certified: it solves the stated model to within the interval. Each tile is judged in the same way, as proven lethal, proven spared, or undecided. Whether that model describes tissue is a question of validation (§ 7).

A Vectorized Checker

The checker above works cell by cell in 256-bit arithmetic, which keeps it easy to audit. A vectorized checker computes the same exact cell integrals on whole arrays of cells in double-precision ball arithmetic [30, 31]. Each quantity is an array of midpoints and an array of radii. Each operation adds a bound on the rounding of its midpoint to the radius, and each radius, itself computed in round-to-nearest, is inflated by 1+γk1+\gamma_k, with γk=ku/(1−ku)\gamma_k=ku/(1-ku), u=2−53u=2^{-53}, and kk at least the number of roundings in its computation. The bounds are rigorous under four assumptions:

  • binary64 arithmetic with round-to-nearest;

  • no flush-to-zero of subnormal numbers, which the checker tests when it runs;

  • sums and matrix products formed by correctly rounded operations in some order, possibly with fused multiply–add, but not by Strassen-type algorithms;

  • the specification’s numbers taken as stored.

The 1/r1/r moments, once per grid column, and the final roots and divisions are computed in Arb. The vectorized checker adds about 390 lines to the trusted base.

On the lesion-map certificate of § 6, it reproduces the bounds of the 256-bit checker. Its η\eta agrees to 2.4×10−62.4 \times10^{-6} relative, its impedance interval differs by about 10−9 Ω10^{-9}\,\Omega, and every tile interval overlaps the 256-bit one. It takes 0.98 s, about a tenth of the solve (Table 1). Its cost grows linearly with the number of cells, while the cost of the solve grows faster, so its share falls to 3.8% at 6.3 million nodal values. The range checker of § 4 is not yet vectorized, and all bounds reported below come from the 256-bit checkers.

SpacingNodal valuesSolveVectorized checkShare of solve256-bit check
0.5 mm157 3602.2 s0.31 s14%5.6 s
0.25 mm506 4769.4 s0.98 s10%17.5 s
0.125 mm1 749 26552.7 s3.35 s6.4%61.0 s
0.0625 mm6 270 646335 s12.6 s3.8%–

Table 1. Cost of Checking Against Solving

Monopolar electrode, degree 3, spacing in the lesion zone; the 0.25 mm grid is the lesion-map certificate. “Nodal values” counts every node of the degree-3 node grid, including inactive ones. MacBook Pro (M5 Pro, 48 GB). Both checks are rigorous.

Bounds Over Parameter Ranges

The electrodes are voltage-controlled, so multiplying both conductivities by one factor leaves uu unchanged. Field quantities therefore depend only on ρ=σb/σt\rho= \sigma_b/\sigma_t. We keep σt\sigma_t at its specified value and let σb=ρσt\sigma_b = \rho\sigma_t range over an interval. Under current control this reduction would fail, because the field then scales with 1/σ1/\sigma.

Proposition 1. Let (vk,Fk)(v_k,F_k) and (vk+1,Fk+1)(v_{k+1},F_{k+1}) be certificates on the same grid, with the same Dirichlet data, computed for ratios ρk<ρk+1\rho_k < \rho_{k+1}. For every s∈[0,1]s \in[0,1], v(s)=(1−s)vk+svk+1v(s) = (1-s)v_k + sv_{k+1} and F(s)=(1−s)Fk+sFk+1F(s) = (1-s)F_k + sF_{k+1} are admissible for every ρ\rho.

Admissibility involves no conductivity, and both conditions are preserved by convex combinations. Theorem 1 therefore holds at each ρ\rho with any ss. We choose s=(ρ−ρk)/Δs = (\rho-\rho_k)/\Delta with Δ=ρk+1−ρk\Delta= \rho_{k+1}-\rho_k, and write η(ρ)\eta(\rho) for the resulting radius. Let σk=σtρk\sigma_k = \sigma_t\rho_k be the blood conductivity of certificate kk and ek=σk∇vk−Fke_k = \sigma_k\nabla v_k - F_k its own residual in blood. Since σ(ρ)−σk=σtΔs\sigma(\rho)-\sigma_k = \sigma_t\Delta s and σ(ρ)−σk+1=−σtΔ(1−s)\sigma(\rho)-\sigma_{k+1} = -\sigma_t\Delta(1-s), the residual in blood is

σ(ρ)∇v(s)−F(s)=(1−s)ek+sek+1+σtΔs(1−s)(∇vk−∇vk+1),(4)\sigma(\rho)\nabla v(s)-F(s) = (1-s)e_k + se_{k+1} + \sigma_t\Delta s(1-s)(\nabla v_k-\nabla v_{k+1}), \tag*{(4)}

and in tissue the last term is absent. So η(ρ)2\eta(\rho)^2 is a quadratic form in (1−s,s,σtΔs(1−s))(1-s,s,\sigma_t\Delta s(1-s)) of inner products among small fields, weighted by 1/σ(ρ)1/\sigma(\rho). In this form, ball evaluation over sub-intervals of ss avoids the cancellation that the naive expansion σ∥∇v∥2−2⟨∇v,F⟩+∥F∥2/σ\sigma\|\nabla v\|^2-2\langle\nabla v,F\rangle+\|F\|^2/\sigma would suffer. The tile centers, which lie in tissue, are ck+s(ck+1−ck)c_k+s(c_{k+1}-c_k) with ck=(σt∇vk+Fk)/2c_k=(\sigma_t\nabla v_k+F_k)/2. The checker accumulates these inner products once. It then evaluates (2) on 64 sub-intervals of ss per segment; ball evaluation is inclusion-monotone, so the bounds hold for every ρ\rho in the interval. The interpolation term shrinks as Δ2\Delta^2, so a modest number of samples brings η\eta back to the level of a single certificate. The approach is close in spirit to majorants for problems with indeterminate data [18], here with blended certificates and rigorous evaluation. Threshold ranges are immediate: a tile is lethal for every threshold in [E1,E2][E_1,E_2] if its lower bound reaches E2E_2. Catheter–tissue gaps change the geometry and are certified one value at a time.

Implementation and Verification

The solver is our own finite-element code, and it is untrusted. The checker trusts no solver, so the bounds do not depend on which one produced the certificate. We wrote our own because standard packages do not output the stream-function current that a certificate needs; § 7.3 certifies a potential computed by FEniCSx with the same checker and the same result (Table 7). Our solver uses degree-pp Lagrange elements for vv and fits ψ\psi by least squares so that FF approximates σ∇v\sigma\nabla v. Where an electrode meets an insulator, the potential behaves like distance\sqrt{\mathit{distance}}. The squared energy error of the first cell then scales with its size; without grading, 99% of η2\eta^2 lay in the two cells at the electrode-shaft junction. Geometric grading down to 1 nm at these points adds only a logarithmic number of grid lines. These singularities belong to the idealized geometry with its sharp corners; real electrodes are rounded. Together with p=3p = 3, the grading brings η\eta from 1.64×10−31.64 \times10^{-3} (degree 1 with weaker grading, 1.3×1051.3 \times10^{5} cells) to 1.3×10−51.3 \times10^{-5} (degree 3, 169×332169 \times332 cells). Relative to ∥c∥\lVert c\rVert the final value is about 1.3×10−41.3 \times10^{-4}.

The test suite checks the following:

  • Certificates from the correct solver at different degrees, resolutions, and voltages give overlapping bounds.

  • Perturbing a single electrode, ring, or ground value, or a stream-function value on an insulating wall, by as little as 10−910^{-9} leads to rejection, also at nodes inside cell edges.

  • A solver that drops the axisymmetric weight, or swaps the two conductivities, still emits admissible certificates. Evaluated with the true physics, the solver’s own claims are refuted or unsupported.

  • With insulating blood, the certified impedance lies inside the window given by Newman’s value 1/(4σta)1/(4\sigma_t a) for a disk on a half-space [32], corrected for the finite ground.

  • On a coarse problem, independent certificates at conductivity ratios not used to build a range certificate lie inside the range bounds.

  • The sampled field of the solver lies inside the certified tile intervals.

The script experiments/paper_checks.py reproduces the remaining checks quoted below. On the full map, independent certificates at ρ=2.2\rho= 2.2 and ρ=3.3\rho= 3.3, which were not used to build the range certificate, lie inside the range bounds on all 768 tiles.

Results

Focal Monopolar Electrode

For the monopolar electrode (L=3L = 3 mm, σt=0.25\sigma_t = 0.25 S/m, σb=0.7\sigma_b = 0.7 S/m, contact) the certified impedance is 93.950 274–93.950 276 Ω\Omega. For this idealized geometry, the value shows how tight the bounds can be, not how accurate the model is. On the 24×3224 \times32 map of 0.25 mm tiles the median relative width of the RMS-field intervals is about 2%. Figure 3 shows the certified lesion at 1000 V and 700 V/cm. It is proven to reach 3.50 mm on the axis and proven not to extend beyond 3.75 mm in any column; for the radius, reached and not beyond coincide at 4.75 mm at tile resolution. Only five tiles on the threshold contour remain undecided. Figure 4 gives the certified extent against voltage for three thresholds. The map covers 6 mm of radius and 8 mm of depth; from about 1500 V the lesion can reach its edge, and those extents are reported as “edge” rather than extrapolated. The certified width-to-depth ratio, about 2.6, is larger than the value of about 2 that Castellvi and Ivorra report for conventional isotropic threshold models of focal monopolar catheters [33]; part of the difference may come from the sharp electrode corner, which inflates the radius at the surface.

Figure 3. Certified lesion of the focal monopolar electrode at 1000 V, shown mirrored about the axis. Orange tiles are proven lethal (Eω≥700E_\omega\ge700 V/cm), hatched tiles are undecided, and all others are proven spared. The background is the solver’s field, which is not certified.

Figure 4. Certified lesion depth and largest radius against voltage. Lines show proven reached (for the depth, on the axis); bands extend to proven not beyond, over every column and row of tiles. The map ends at 8 mm depth and 6 mm radius.

Conductivity and Threshold Uncertainty

Blood conductivities of 0.6–0.75 S/m and myocardial conductivities of 0.2–0.35 S/m give ratios within [1.7,3.75][1.7, 3.75]. Seventeen certificates cover this interval with ηmax⁡=1.52×10−5\eta_{\max} = 1.52 \times10^{-5}. The median and worst relative widths of the tile intervals are 8% and 27%, now including the true variation over ρ\rho. With five certificates, interpolation error raised ηmax⁡\eta_{\max} to 1.4×10−41.4 \times10^{-4} and widened the proven depth band to 3.0–4.5 mm. That widening came from interpolation, not from the conductivity. Point certificates at ρ=1.7\rho= 1.7, 2.2, 3.3, and 3.75 each prove a depth between 3.50 mm and 3.75 mm. Table 2 shows that, within this range, the conductivity barely moves the lesion, while a ±100\pm100 V/cm uncertainty in the threshold widens the proven band considerably.

VoltageThreshold (V/cm)Depth, nominalDepth, all ρ\rhoRadius, nominalRadius, all ρ\rho
500 V7002.00–2.002.00–2.003.25–3.253.00–3.25
500 V600–8001.75–2.501.75–2.503.00–3.502.75–3.50
1000 V7003.50–3.753.50–3.754.75–4.754.75–5.00
1000 V600–8003.25–4.003.25–4.254.50–5.254.25–5.25
1500 V7004.75–5.004.50–5.006.00–edge5.75–edge
1500 V600–8004.25–5.504.25–5.505.50–edge5.50–edge
2000 V7005.50–6.005.50–6.006.00–edge6.00–edge
2000 V600–8005.00–6.505.00–6.756.00–edge6.00–edge

Table 2. Certified Lesion of the Monopolar Electrode for the Nominal and for All Conductivity Ratios. Entries are proven reached to proven not beyond (mm), for the nominal ratio ρ=2.8\rho= 2.8 and for every ρ∈[1.7,3.75]\rho\in[1.7, 3.75]; “edge” means the lesion may reach the edge of the map.

Comparison With a Grid Convergence Study

A practitioner would verify the monopolar solution by solving on systematically refined grids and applying Richardson extrapolation with the grid convergence index (GCI) [6, 34]. We did this with the same solver at degree 2, on two families of four grids with refinement ratio 2 and spacing from 1 mm to 0.125 mm in the lesion zone. One family is uniform near the electrode, as a typical user would mesh it; the other is graded toward the electrode corners (the tip edge and the electrode-shaft junction) as in § 5. The graded family scales the grading slope and the far-field growth with the spacing but keeps its smallest spacing at 1 nm. Both families keep grid lines at the geometry and tile boundaries. The grids are therefore not geometrically similar near the corners, and the observed orders there should be read with this in mind. On every grid we recorded what the finite-element solution alone gives (the impedance from the field energy and the tile RMS field from the potential gradient) and the certified interval for the same quantity. The GCI uses the three finest grids and a safety factor of 1.25. The convergence ratio RR classifies each triplet; the GCI is valid only for monotone convergence, 0<R<10 < R < 1 [35]. The reference is a degree-3 certificate on the finest graded grid. Table 3 summarizes the results for the impedance and two tiles at 1000 V: one touching the electrode edge, and one on the axis at the lesion boundary.

QuantityGridsFinite element, finest gridGCI: RR, order, bandCertified, finest gridReference
Impedance (Ω\Omega)Uniform93.81630.88, 0.19, 92.60–95.0393.8163–94.081893.9502744–93.9502758
Graded93.9502700.06, 3.96, 93.950264–93.95027593.950269–93.950280
Edge tile (V/cm)Uniform57085.2, n/a, 5690–57273306–80275664.1–5674.7
Graded5669.40.06, 4.00, 5669.4025–5669.40385654.7–5684.1
Axis tile (V/cm)Uniform693.61.14, n/a, 691.6–695.60–8522676.1–710.9
Graded693.50.07, 3.91, 693.4809–693.4812644.7–742.2

Table 3. Grid Convergence Index Versus Certified Intervals for the Monopolar Electrode. Degree 2, four grids per family with spacing 1–0.125 mm; tile RMS fields at 1000 V. The GCI uses the three finest grids; R>1R > 1 means the triplet diverges, the GCI is not valid, and no order is given. Reference: a degree-3 certificate on the finest graded grid, hence tighter than the interval of Section 6. Edge tile: 1.25 ≤r≤\le r \le 1.5 mm, −0.25≤z≤0-0.25 \le z \le0 mm; axis tile: 0≤r≤0 \le r \le 0.25 mm, −3.75≤z≤−3.5-3.75 \le z \le-3.5 mm.

On the graded grids, both methods work. Every triplet converges at an observed order of about 4, and every GCI band is consistent with the reference. For the impedance the certified interval is as narrow as the GCI band. Even here, the extrapolated impedance, 93.950 273 8 Ω, lies just below the certified reference interval; the difference of 6×10−76 \times10^{-7} Ω is irrelevant in practice but shows that extrapolation carries no guarantee. For the two tiles the certified intervals, about ±0.3% and ±7%, are far wider than the GCI bands (about ±10−510^{-5}%), because every tile bound uses the global η\eta. Where the grid is good, a careful mesh study is sharper for local quantities, although it remains an estimate.

The uniform grids behave differently. The impedance still converges, but at an observed order of 0.19, far from the asymptotic range, so the GCI band is ±1.3%; the certificate from the finest grid alone is about nine times narrower and proven. At the tile touching the electrode edge, refinement makes the change larger rather than smaller (R=5.2R = 5.2): the triplet diverges and the GCI is not valid. Applied regardless, it reports ±0.33%\pm0.33\% around a value 0.6% above the upper end of the reference interval, a band that excludes the true value. The certified interval on the same grid is wide but contains the reference, and its width indicates that the grid is inadequate near the edge. On the axis tile the triplet also diverges, slowly (R=1.14R = 1.14), and the certified interval is too wide to be useful.

A mesh study checks convergence, not correctness. It compares a solver only with itself, so even an expert running it cannot tell whether the value it converges to is the one the specification defines. We repeated it on the graded grids with the two solvers of § 5 that model the wrong physics. Both converge cleanly, at an observed order of about 4, and the GCI gives tight bands around their own impedances, 136.806 Ω136.806\,\Omega with the conductivities swapped and 0.393 Ω0.393\,\Omega without the axisymmetric weight, against a true 93.950 Ω93.950\,\Omega. A mesh study accepts both. Checked against the specification on the finest grid, the claim of the solver without the axisymmetric weight lies outside its certified interval, 40.340.3–145.4 Ω145.4\,\Omega, and is refuted. The solver with swapped conductivities receives 88.688.6–220.9 Ω220.9\,\Omega, which contains its claim but cannot prove it to 1%, so the claim is unsupported; on the same grid the correct solver’s interval is 10710^7 times narrower. Code verification with manufactured solutions or a review of the setup could also catch such errors, but as separate work rather than as part of every run, and the manufactured source must be derived from an independent statement of the physics.

The two methods are therefore complementary for conventional solvers. A mesh study estimates the discretization error and is often sharp in the asymptotic range. The certificate gives a guaranteed bracket for the specified problem: it falsifies a GCI band disjoint from it, corroborates one it tightly contains, and says nothing either way when it is wide. It holds on every grid from a single solve. Its present weakness is that local intervals are wider than a converged GCI band. The script experiments/gci_comparison.py reproduces Table 3 and the wrong-physics comparison.

Bipolar Catheter and Comparison With Meckes et al.

Meckes et al. modeled a two-electrode catheter with 2 mm electrodes 4 mm apart. They used a lethal threshold of 268 V/cm and compared their predictions with in-vivo lesion depths [1]. Their conductivities are not given in the main text. Several geometric details are not stated precisely: the catheter orientation, the electrode radius, and whether the spacing is measured edge to edge. We assume a catheter perpendicular to the tissue (axisymmetric), an electrode radius of 1.25 mm, a 4 mm insulating spacer, a floating generator, and no temperature dependence. We replace the missing conductivities by a range. At contact, 17 certificates cover ρ∈[1.7,3.75]\rho\in[1.7, 3.75] with ηmax⁡=1.7×10−5\eta_{\max} = 1.7 \times10^{-5}. Gaps of 1–4 mm are certified at the nominal ratio. With Thomson’s bound, the tip-to-ring impedance at contact is certified as 172.6044 Ω\Omega at the nominal ratio, tight to the last digit shown. The median and worst relative widths of the range tile intervals are 18% and 68%.

Table 4 and Figure 5 show that the certified depths overlap the in-vivo data at both voltages. These data come from other studies cited by Meckes et al., at 500 V and 2200 V, with different catheters and waveforms, and their intervals are wide. The overlap is therefore a plausibility check and does not validate the model. For the gap, Meckes et al. report that 1–2 mm roughly doubles the voltage needed for 3 mm; the strong effect of electrode-tissue proximity on PFA lesions has also been studied in tissue [36]. Table 5 agrees at 2 mm, where the factor lies between 1.77 and 2.47, and gives a smaller factor at 1 mm, between 1.23 and 1.63. At low voltage our lesions are deeper than in their model. We reach 3 mm with at most 550 V, whereas Meckes et al. need 700 V, and 1 mm with at most 175 V against their 300 V. At 1200 V and a 4 mm gap we certify 1.0–1.5 mm where they report a negligible lesion.

VoltageCertified (mm)Meckes model (mm)In vivo (mm), as reported by Meckes et al.
500 V2.75–3.002.282.25 ±0.85\pm0.85 (1.35–3.14)
2200 V5.75–7.506.166.5 ±1.7\pm1.7 (3.8–9.2)

Table 4. Certified Depth of the Bipolar Catheter Compared With Meckes et al. Bipolar catheter at contact, 268 V/cm. Certified: proven reached to proven not beyond, for every ρ∈[1.7,3.75]\rho\in[1.7, 3.75]. In vivo: as reported by Meckes et al., with 95% confidence intervals in parentheses.

Gap1 mm deep3 mm deep
0 mm150 / 175 V475 / 550 V
1 mm225 / 275 V675 / 775 V
2 mm400 / 450 V975 / 1175 V
3 mm650 / 750 V1350 / 1750 V
4 mm950 / 1175 V1800 / >2500 V

Table 5. Voltage Needed for a Given Depth With a Blood Gap Under the Tip. Nominal ratio; largest voltage proven insufficient / smallest voltage proven sufficient, on a 25 V grid.

Figure 5. Bipolar catheter: certified depth against voltage at contact (every ρ∈[1.7,3.75]\rho\in[1.7,3.75]) and at a 2 mm gap (nominal ρ\rho). Lines show proven reached and bands extend to proven not beyond; bands reaching the 8 mm map edge are open-ended. At high voltage the threshold is met deep in tissue, where the field decays slowly, so several tiles lie within their certified uncertainty of the threshold and the band widens. The squares are Meckes et al.’s model predictions and the circles the in-vivo depths they cite, with ±\pm one standard deviation.

The certificates narrow down where this difference can come from. They prove that it is not numerical error in our computation; a converged mesh study would indicate the same without proof. They also exclude every constant conductivity ratio [1.7,3.75]. Outside that interval the depth does depend on the ratio (Table 6). A lower ratio gives a shallower lesion. Within constant-conductivity models, ρ=0.7\rho= 0.7 reproduces two of Meckes et al.’s figures: the certified depth at 500 V is 2.25–2.50 mm, which contains their 2.28 mm, and 700 V is proven sufficient for 3 mm, their reported minimum. This ratio means myocardium more conductive than blood, 1.0 S/m at σb=0.7\sigma_b = 0.7 S/m, about three times the top of the 0.2–0.35 S/m range, which is at the edge of plausibility. Their model includes field- and temperature-dependent conductivity, but such a model raises the conductivity most near the electrode, where the field is strongest, and in sequential electroporation models this can enlarge rather than shrink the electroporated region [37]. Whether their nonlinearity acts like a lower effective ratio cannot be decided without their model. Even at ρ=0.7\rho= 0.7, the 1 mm threshold differs. We need at most 200 V against their 300 V.

ρ=σb/σt\rho= \sigma_b/\sigma_tDepth at 500 V (mm)3 mm needs (V)1 mm needs (V)
0.72.25–2.50650 / 700175 / 200
1.02.50–2.50575 / 650150 / 175
2.82.75–3.00475 / 550150 / 175
7.03.00–3.00425 / 500125 / 175
103.00–3.25425 / 500125 / 175
Meckes et al.2.28700300

Table 6. Bipolar Catheter: Point Certificates Inside and Outside the Conductivity Range. Contact, 268 V/cm; voltages as proven insufficient / proven sufficient.

Other candidates are further differences between the two models:

  • Catheter orientation (a perpendicular tip drives current straight into the wall; a catheter lying along it does not).

  • Electrode radius and spacing.

  • Boundary conditions.

  • Field- and temperature-dependent conductivity, as discussed above.

  • The lesion criterion (tile RMS here, pointwise there, a difference of about one tile where the field decreases monotonically).

  • Numerical error in the reference model, which certificates of our own model cannot exclude.

Deciding among these needs their exact model and, for field-dependent conductivity and orientation, the extensions of § 7. A disagreement that survives the certificates points at the physics or at the other computation.

Discussion

What a Certificate Establishes

Every bound is a theorem about the stated model: geometry, conductivities, and boundary conditions as written in the specification. Provided the checker and the specification are correct, the interval contains the stated model’s true value whatever a solver returns, and a wrong solver is rejected, refuted, or unsupported (§ 3). Trust therefore rests on the checker and the specification. The checker itself needs code verification, which the test suite provides in part. The theorems used here are short. Formalizing them in a proof assistant, together with a verified implementation of the checker, would leave the specification and a proof kernel as the trusted base.

A tight certificate verifies the computed solution of that run. In the monopolar example, η\eta is about 1.3×10−41.3 \times10^{-4} of ∥c∥\lVert c\rVert. The computed current density is therefore proven to lie within 0.013% of the exact solution of the stated problem in the energy norm, although the exact solution is never known, just as a two-sided bracket proves digits of 2\sqrt{2}. A verified solver would give the same conclusion for this input. The certificate reaches it for a single run, a far smaller task than verifying a solver for all inputs. A wide bound is inconclusive: it states only that this output is not proven accurate to a useful precision.

A wrong specification, with wrong units, geometry, or conductivity (use error in V&V 40 terms), is certified as faithfully as a right one. A wrong model is a separate failure, addressed by validation and in part by certifying over parameter ranges, which replaces “for our guessed conductivity” by “for every conductivity in this range”. The lethal threshold is itself contested. Petras et al. find that no single threshold fits both depth and width, with roughly 270–500 V/cm for depth and 790–1000 V/cm for width [2].

Use in Credibility Assessment

Certificates provide rigorous calculation verification for every run in the context of use, rather than for a representative case. For patient-specific clinical tools, the highest example level of discretization-error evidence proposed by Galappatthige et al. is a convergence analysis performed automatically on each new patient, with the error estimated and used in processing the results [38]; a certificate replaces that estimate by a proven bound. The inner envelope supports efficacy claims about the target at tile resolution, whereas safety claims about collateral structures need pointwise bounds. Certificates address the numerical part of credibility, on which the rest depends. A comparison with data is informative only once the numerical error is known to be small. Validation, the applicability of the validation evidence, and uncertainty quantification remain separate tasks, and they carry most of the weight in medical device submissions [39, 40, 38]. Certificates over parameter ranges also contribute to the last.

A point certificate costs about a tenth of a solve to check with the vectorized checker (§ 3.4), so a check can accompany every run. The range certificate over 17 ratios takes 2.6 min of solves and 7.8 min of checking, because the range checker still works cell by cell in 256-bit arithmetic. Like the point check, it is a single pass of local integrals with no system to solve.

Other Solvers and Learned Surrogates

In principle, any solver that outputs fields can participate. A conventional finite-element solution provides an admissible potential, and a flux reconstruction provides the current, for instance by local equilibration [15, 16] or by the least-squares stream-function fit our solver uses. A finite-volume solution provides the current, and a potential reconstruction provides the rest.

As a check of independence, the potential of the monopolar lesion-map problem was also computed by FEniCSx [41], with its own assembly, quadrature, and linear solver. It used the same grid and the same degree-3 tensor elements, and the stream-function fit, used as an untrusted adapter, supplied the current. Table 7 shows that the unchanged checker gives the same bounds as for our solver. The two potentials differ by at most 7×10−9 V7 \times10^{-9}\ \mathrm{V}, and FEniCSx’s own impedance equals the certified lower bound to the digits shown. This tests an independent implementation on the same discretization. Certificates for solvers on their own meshes are left to follow-up work.

For PFA, neural networks already predict the ablated area, validated against experimental data from the literature [42]. Their error comes from approximation and training, so Richardson extrapolation and the GCI do not apply, and test-set and conformal guarantees hold in probability over an input distribution rather than for a given output. A surrogate output that is made admissible can be certified like any other, and the bound remains valid outside the training distribution. Bounds of this functional a posteriori type have been proposed for neural solutions [23]; the independent checker, the rigorous arithmetic, and the written specification are what this paper adds. An output can be made admissible by post-processing, which without Galerkin orthogonality needs a global correction [25], or by construction, with a stream function whose wall constants are imposed exactly and a potential lifted to the Dirichlet data. Neither route is demonstrated here.

Limitations and Extensions

The model is axisymmetric, linear, isotropic, and quasi-static, with constant conductivities. It has no frequency dispersion, no temperature dependence, and no field dependence; the last would also break the linearity in VV that lets one certificate serve all voltages. Electrodes are voltage-controlled flat-tipped cylinders with sharp corners, and the corner touching the tissue inflates the lesion radius at the surface. The monopolar return is a grounded box 50 mm away rather than a skin patch.

Real lesions depend on more than this model contains. The lethal threshold depends on the waveform and on whether lesions are read acutely or chronically [43, 44]. The tissue has finite wall thickness, so the clinical question is transmurality, and it contains fat, scar, and trabeculation. Electrodes indent the tissue under contact force, repeated high-voltage trains heat tissue and can form gas at the electrodes, and clinical multielectrode catheters produce three-dimensional fields. The perpendicular tip and tip-to-ring configurations here are stylized test cases for the method.

The method has limits of its own. Lesion criteria use tile RMS fields. Gaps are certified at sampled values only. The comparison with Meckes et al. rests on our assumptions about their geometry. Bands widen with depth because every tile bound uses the global η\eta, so reducing η\eta would tighten them and smaller tiles would not. Pointwise bounds could be added, since in each homogeneous region ∣∇u∣2|\nabla u|^2 is subharmonic and its value at a point is at most its mean over any ball around the point that lies in that region.

The orientation question of § 6 needs a three-dimensional model. Nothing in the certificate depends on axisymmetry. On a tetrahedral mesh, the potential is a continuous piecewise polynomial with the electrode and ground values. The current is best supplied as the curl of a Nédélec vector potential, the three-dimensional analog of the stream function used here, because a Raviart–Thomas field with floating-point coefficients is not exactly divergence-free. Alternatively, a small residual divergence can be bounded by a term involving a Friedrichs constant [17]. The checker then integrates products of polynomials over tetrahedra in closed form, and the range method of § 4 carries over unchanged. A polygonal mesh does not represent a curved electrode exactly, so the specification would state the meshed geometry, or the geometric error would need a separate bound. Electrode edges become singular lines that need anisotropic grading, and meshes of a hundred thousand to a million elements call for a vectorized rigorous checker.

Field-dependent conductivity, the other candidate in § 6, needs no third dimension. With a conductivity that increases with ∣∇u∣|\nabla u|, the problem still minimizes a convex energy, and a convex-duality analog of Theorem 1 gives certificates of the same kind [17]. For semilinear problems, equilibrated fluxes already certify existence and outputs [26].

Conclusion

When simulations run without an expert checking each result, the check has to come with the result. We have shown that each run can be verified by its own certificate. The solver supplies an admissible potential and current, and a small checker, independent of it, turns them into guaranteed bounds against a written specification of the physics. Its vectorized version checks a single certificate at about a tenth of the cost of the solve. For PFA, the bounds certify lesion maps at the 0.25 mm tile resolution near the threshold contour at moderate voltages, with bands that widen with depth. They extend to whole ranges of uncertain tissue conductivity, they remain valid on grids where a standard mesh study breaks down, and they rule out our own numerical error as the cause of a disagreement with another model. For PFA device modeling, this means that each predicted lesion, for a given catheter, voltage, and range of tissue conductivity, comes with a proven numerical error, so that a comparison with tissue data tests the model alone. For conventional solvers, mesh studies and certificates are complementary: a mesh study estimates the error, and a certificate proves a bound that can falsify the estimate. Learned surrogates have no grid to refine, so a guarantee for a given output needs an a posteriori bound of this kind. The method is limited today to linear, axisymmetric conduction. Localized bounds, three-dimensional geometry, and certificates from other solvers and from learned surrogates are the next steps.

Code Availability

The solver, checker, tests, and the scripts that reproduce every number in this paper are available at https://github.com/kostisdov/certsim (tag v1.0-preprint) under the Apache License 2.0.

References

  1. [1]D. Meckes, M. Emami, I. Fong, D. H. Lau, and P. Sanders, “Pulsed-field ablation: Computational modeling of electric fields for lesion depth analysis,” Heart Rhythm O2, vol. 3, no. 4, pp. 433–440, 2022, doi: 10.1016/j.hroo.2022.05.009.DOI
  2. [2]A. Petras et al., “Is a single lethal electric field threshold sufficient to characterize the lesion size in computational modeling of cardiac pulsed-field ablation?” Heart Rhythm O2, vol. 6, no. 5, pp. 671–677, 2025, doi: 10.1016/j.hroo.2025.02.014.DOI
  3. [3]A. Anić, A. Ivorra, B. Koop, N. J. Kirkpatrick, and J. Koruth, “Pulsed field ablation: From concepts to technology,” Europace, vol. 28, no. Suppl. 2, pp. ii18–ii31, 2026, doi: 10.1093/europace/euag205.DOI
  4. [4]American Society of Mechanical Engineers, Assessing Credibility of Computational Modeling Through Verification and Validation: Application to Medical Devices, ASME Standard V&V 40-2018, New York, NY, USA, 2018.
  5. [5]U.S. Food and Drug Administration, “Assessing the credibility of computational modeling and simulation in medical device submissions,” Guidance for Industry and FDA Staff, Docket FDA-2021-D-0980, Nov. 2023.
  6. [6]P. J. Roache, Verification and Validation in Computational Science and Engineering. Albuquerque, NM, USA: Hermosa, 1998.
  7. [7]W. L. Oberkampf and C. J. Roy, Verification and Validation in Scientific Computing. Cambridge, U.K.: Cambridge Univ. Press, 2010, doi: 10.1017/CBO9780511760396.DOI
  8. [8]L. Lu, P. Jin, G. Pang, Z. Zhang, and G. E. Karniadakis, “Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators,” Nat. Mach. Intell., vol. 3, no. 3, pp. 218–229, 2021, doi: 10.1038/s42256-021-00302-5.DOI
  9. [9]G. C. Necula, “Proof-carrying code,” in Proc. 24th ACM SIGPLAN-SIGACT Symp. Princ. Program. Lang. (POPL), Paris, France, 1997, pp. 106–119, doi: 10.1145/263699.263712.DOI
  10. [10]R. M. McConnell, K. Mehlhorn, S. Näher, and P. Schweitzer, “Certifying algorithms,” Comput. Sci. Rev., vol. 5, no. 2, pp. 119–161, 2011, doi: 10.1016/j.cosrev.2010.09.009.DOI
  11. [11]W. Prager and J. L. Synge, “Approximations in elasticity based on the concept of function space,” Quart. Appl. Math., vol. 5, no. 3, pp. 241–269, 1947, doi: 10.1090/qam/25902.DOI
  12. [12]J. L. Synge, The Hypercircle in Mathematical Physics. Cambridge, U.K.: Cambridge Univ. Press, 1957.
  13. [13]G. Pólya and G. Szegő, Isoperimetric Inequalities in Mathematical Physics. Princeton, NJ, USA: Princeton Univ. Press, 1951, doi: 10.1515/9781400882663.
  14. [14]P. Ladevèze and D. Leguillon, “Error estimate procedure in the finite element method and applications,” SIAM J. Numer. Anal., vol. 20, no. 3, pp. 485–509, 1983, doi: 10.1137/0720033.
  15. [15]D. Braess, V. Pillwein, and J. Schöberl, “Equilibrated residual error estimates are p-robust,” Comput. Methods Appl. Mech. Eng., vol. 198, no. 13–14, pp. 1189–1197, 2009, doi: 10.1016/j.cma.2008.12.010.DOI
  16. [16]A. Ern and M. Vohralík, “Polynomial-degree-robust a posteriori estimates in a unified setting for conforming, nonconforming, discontinuous Galerkin, and mixed discretizations,” SIAM J. Numer. Anal., vol. 53, no. 2, pp. 1058–1081, 2015, doi: 10.1137/130950100.DOI
  17. [17]S. Repin, A Posteriori Estimates for Partial Differential Equations. Berlin, Germany: De Gruyter, 2008, doi: 10.1515/9783110203042.DOI
  18. [18]O. Mali, P. Neittaanmäki, and S. Repin, Accuracy Verification Methods: Theory and Algorithms. Dordrecht, The Netherlands: Springer, 2014, doi: 10.1007/978-94-007-7581-7.DOI
  19. [19]X. Liu and S. Oishi, “Verified eigenvalue evaluation for the Laplacian over polygonal domains of arbitrary shape,” SIAM J. Numer. Anal., vol. 51, no. 3, pp. 1634–1654, 2013, doi: 10.1137/120878446.DOI
  20. [20]X. Liu, M. T. Nakao, C. You, and S. Oishi, “Explicit a posteriori and a priori error estimation for the finite element solution of Stokes equations,” Jpn. J. Ind. Appl. Math., vol. 38, no. 2, pp. 545–559, 2021, doi: 10.1007/s13160-020-00449-5.DOI
  21. [21]M. T. Nakao, M. Plum, and Y. Watanabe, Numerical Verification Methods and Computer-Assisted Proofs for Partial Differential Equations. Singapore: Springer, 2019, doi: 10.1007/978-981-13-7669-6.DOI
  22. [22]G. Rozza, D. B. P. Huynh, and A. T. Patera, “Reduced basis approximation and a posteriori error estimation for affinely parametrized elliptic coercive partial differential equations,” Arch. Comput. Methods Eng., vol. 15, no. 3, pp. 229–275, 2008, doi: 10.1007/s11831-008-9019-9.DOI
  23. [23]V. Fanaskov, A. Rudikov, and I. Oseledets, “Neural functional a posteriori error estimates,” arXiv:2402.05585, 2024.arxiv.org/abs/2402.05585
  24. [24]M. Figel and C. Merdon, “Local equilibration error estimators for guaranteed error control in adaptive stochastic higher-order Galerkin finite element methods,” SIAM/ASA J. Uncertain. Quantif., vol. 4, no. 1, pp. 1372–1397, 2016, doi: 10.1137/15M102188X.
  25. [25]J. Papež, Z. Strakoš, and M. Vohralík, “Estimating and localizing the algebraic and total numerical errors using flux reconstructions,” Numer. Math., vol. 138, no. 3, pp. 681–721, 2018, doi: 10.1007/s00211-017-0915-5.DOI
  26. [26]H. Ishizaka, “Low-order CR–RT equilibrated-flux certification for semilinear problems on anisotropic meshes,” arXiv:2602.01636, 2026.arxiv.org/abs/2602.01636
  27. [27]K. K. H. Cheung, A. Gleixner, and D. E. Steffy, “Verifying integer programming results,” in Integer Programming and Combinatorial Optimization (IPCO), Lecture Notes in Computer Science. Cham, Switzerland: Springer, 2017, pp. 148–160, doi: 10.1007/978-3-319-59250-3_13.DOI
  28. [28]F. Johansson, “Arb: Efficient arbitrary-precision midpoint-radius interval arithmetic,” IEEE Trans. Comput., vol. 66, no. 8, pp. 1281–1292, 2017, doi: 10.1109/TC.2017.2690633.DOI
  29. [29]M. Paraschivoiu, J. Peraire, and A. T. Patera, “A posteriori finite element bounds for linear-functional outputs of elliptic partial differential equations,” Comput. Methods Appl. Mech. Eng., vol. 150, no. 1–4, pp. 289–312, 1997, doi: 10.1016/S0045-7825(97)00086-8.
  30. [30]S. M. Rump, “Fast and parallel interval arithmetic,” BIT Numer. Math., vol. 39, no. 3, pp. 534–554, 1999, doi: 10.1023/A:1022374804152.DOI
  31. [31]S. M. Rump, “Verification methods: Rigorous results using floating-point arithmetic,” Acta Numer., vol. 19, pp. 287–449, 2010, doi: 10.1017/S096249291000005X.DOI
  32. [32]J. Newman, “Resistance for flow of current to a disk,” J. Electrochem. Soc., vol. 113, no. 5, p. 501, 1966, doi: 10.1149/1.2424003.DOI
  33. [33]Q. Castellví and A. Ivorra, “Computational multiscale modeling of pulsed field ablation considering conductivity and damage anisotropy reveals deep lesion morphologies,” Int. J. Numer. Methods Biomed. Eng., vol. 41, no. 8, Art. no. e70077, 2025, doi: 10.1002/cnm.70077.DOI
  34. [34]I. B. Celik, U. Ghia, P. J. Roache, C. J. Freitas, H. Coleman, and P. E. Raad, “Procedure for estimation and reporting of uncertainty due to discretization in CFD applications,” J. Fluids Eng., vol. 130, no. 7, Art. no. 078001, 2008, doi: 10.1115/1.2960953.DOI
  35. [35]F. Stern, R. V. Wilson, H. W. Coleman, and E. G. Paterson, “Comprehensive approach to verification and validation of CFD simulations—Part 1: Methodology and procedures,” J. Fluids Eng., vol. 123, no. 4, pp. 793–802, 2001, doi: 10.1115/1.1412235.DOI
  36. [36]B. Howard et al., “Effects of electrode-tissue proximity on cardiac lesion formation using pulsed field ablation,” Circ. Arrhythm. Electrophysiol., vol. 15, no. 10, Art. no. e011110, 2022, doi: 10.1161/CIRCEP.122.011110.DOI
  37. [37]D. Šel, D. Cukjati, D. Batuskaite, T. S. S̆livnik, L. M. Mir, and D. Miklavčič, “Sequential finite element model of tissue electroporation,” IEEE Trans. Biomed. Eng., vol. 52, no. 5, pp. 816–827, 2005, doi: 10.1109/TBME.2005.845212.DOI
  38. [38]S. Galappaththige, R. A. Gray, C. J. Mendonca Costa, S. Niederer, and P. Pathmanathan, “Credibility assessment of patient-specific computational modeling using patient-specific cardiac modeling as an exemplar,” PLoS Comput. Biol., vol. 18, no. 10, Art. no. e1010541, 2022, doi: 10.1371/journal.pcbi.1010541.DOI
  39. [39]P. Pathmanathan, R. A. Gray, V. J. Romero, and T. M. Morrison, “Applicability analysis of validation evidence for biomedical computational models,” J. Verif. Valid. Uncertainty Quantif., vol. 2, no. 2, Art. no. 021005, 2017, doi: 10.1115/1.4037671.DOI
  40. [40]M. Viceconti, F. Pappalardo, B. Rodriguez, M. Horner, J. Bischoff, and F. Musuamba Tshinanu, “In silico trials: Verification, validation and uncertainty quantification of predictive models used in the regulatory evaluation of biomedical products,” Methods, vol. 185, pp. 120–127, 2021, doi: 10.1016/j.ymeth.2020.01.011.DOI
  41. [41]I. A. Baratta, J. P. Dean, J. S. Dokken, M. Habera, J. S. Hale, C. N. Richardson, M. E. Rognes, M. W. Scroggs, N. Sime, and G. N. Wells, “DOLFINx: The next generation FEniCS problem solving environment,” Zenodo, 2023, doi: 10.5281/zenodo.10447666.
  42. [42]R. Crusi, N. Colistra, F. Camera, G. Monti, M. S. Zappatore, C. Merla, and L. Tarricone, “Cardiac pulsed-field ablation: Deep learning solutions for multi-parameter predictions,” IEEE J. Electromagn. RF Microw. Med. Biol., vol. 9, no. 4, pp. 445–454, 2025, doi: 10.1109/JERM.2025.3577268.DOI
  43. [43]T. Kotnik, L. Rems, M. Tarek, and D. Miklavčič, “Membrane electroporation and electropermeabilization: Mechanisms and models,” Annu. Rev. Biophys., vol. 48, no. 1, pp. 63–91, 2019, doi: 10.1146/annurev-biophys-052118-115451.DOI
  44. [44]A. Verma, S. J. Asirvatham, T. Deneke, Q. Castellvi, and R. E. Neal II, “Primer on pulsed electrical field ablation: Understanding the benefits and limitations,” Circ. Arrhythm. Electrophysiol., vol. 14, no. 9, 2021, doi: 10.1161/CIRCEP.121.010086.DOI

Paper details

vlnc:W_x5ga34t9
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

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

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents