acceptodds
Under review as a conference paper at ICLR 2027

Enriched Holomorphic Neural Networks for 2D and 3D Laplace Problems

Abstract

Boundary-value problems for the Laplace equation develop singular behaviour near corners, edges and junctions that smooth approximators resolve poorly. We combine holomorphic neural networks with singular functions derived from the local geometry and boundary conditions. The representation is harmonic by construction, so no interior residual is needed: the network fits the smooth remainder from boundary data alone, and the enrichment terms capture the singular behaviour. Singular exponents are trainable, so the construction also applies when the local expansion is not known in advance. We extend the approach to three dimensions using independent networks evaluated along complex null directions. The enriched model is the most accurate tested method on every problem. Its margin is several orders of magnitude where the enrichment captures the singularity exactly, and a factor of about 1.7–2.1 over the next-best method on a CdZnTe radiation detector with many singular sites, in both two and three dimensions.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Physics-Informed Neural Network, Laplace Equation, Holomorphic Networks, Harmonic functions, Mesh-free Methods

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_NkyfOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes enriched holomorphic neural networks for 2D and 3D Laplace problems with singularities. The method combines holomorphic networks with analytical singular functions and trainable singular exponents. It extends to 3D using independent networks along complex null directions. Experiments cover the closed-form, series reference, and application problems. Ablation studies examine singularity enrichment and training settings. The results show substantial accuracy improvements over tested baselines.

Critical strengths
  1. The combination of the singularity enrichment dictionary, trainable singular exponents and the extension to three dimensions helps improve the general framework of solving Laplace problems with singularities. It is a theoretical advantage that the model satisfies the Laplace equation by construction.

  2. The proposed methodology is evaluated across multiple experimental settings, including 2D and 3D singularity benchmarks, the randomized setups (24 instances, 72 paired runs). Experiments also show that the enrichment terms achieve more accurate solutions. The ablation on training budget and ladder depth also help explain the performance improvement.

Critical weaknesses
  1. The largest improvements are the ones that exact solutions can be represented directly by the supplied singular modes. In the CZT detector problem, the gains become substantially smaller, only 1.7\times+ over the next-best baselines. More evaluations of e.g. error variance, convergence rate across more random initializations, would strengthen the evidence of robustness.
Questions for authors
  1. Beyond the existing ablation studies, how does the solution error scale with network size and the number of singular enrichment terms across the 2D and 3D benchmarks? Does the same trend hold for problems whose exact solutions cannot be represented directly by the supplied singular modes?
2 comments

We thank the reviewer for their feedback and aim to address the weakness and question highlighted below.

Weakness 1: We agree that the largest improvements occur on problems whose solutions are supplied enrichment terms. We therefore also assessed problems where the network must represent a remainder. On Motz's problem and the cracked beam, adding the terms reduces the error by factors of 2.2×1042.2\times10^4 and 7.4×1037.4\times10^3, respectively (Table R2). In 3D, on uAu_A, uBu_B and extruded Motz, the error decreases by factors of 75–250 compared with the same model without the terms, and by a factor of 36 on the 3D detector (Table R2). On the 2D detector, the improvement is smaller, at a factor of 1.7, as discussed in Section 5.2. These results show that enrichment terms are also useful when they do not represent the entire solution.

To assess variance, we ran uAu_A and uBu_B with 20 seeds per model, each with three restarts. All 80 seed results, including failures, are retained in the medians and comparisons; no seed is dropped. With enrichment, the error lies between 9.6×10−59.6\times10^{-5} and 2.2×10−42.2\times10^{-4} on 39 of 40 seeds. The remaining run (uBu_B, seed 8) has relative error 0.28 and boundary loss 0.16, compared with a boundary loss of at most 1.9×10−61.9\times10^{-6} for the other runs. Without enrichment, 10 of 20 seeds on each problem finish with errors above 10−210^{-2}, including 4 of 20 per problem with errors above 1. We believe these runs diverged, as their relative L2L^2 errors are worse than predicting zero. On the 3D detector, the errors from our three seeds (1.1×10−21.1\times10^{-2}–1.5×10−21.5\times10^{-2}) do not overlap with those of the next-best methods (2.3×10−22.3\times10^{-2} and above). These experiments assess final accuracy across seeds; they do not measure convergence rates during training.

Question 1: We ran additional scaling experiments on problems whose solutions are not sums of the supplied enrichment terms and observe the same trend. In 2D, we tested Motz's problem and the cracked beam with widths from 4 to 64 and 0 to 8 enrichment terms. Increasing network size alone does not improve accuracy: without enrichment, the error remains between 1.8×10−31.8\times10^{-3} and 4.2×10−34.2\times10^{-3} as the parameter count increases from 78 to 8,718. Adding the first term reduces the error by a factor of 90–300, and the second gives a further reduction by a factor of 6–450. With four terms, the error decreases as width increases. For example, on Motz's problem it decreases from 8.5×10−78.5\times10^{-7} at width 4 to 6.0×10−86.0\times10^{-8} at width 64. Using eight terms gives less accurate results than using four, with errors between 1.5×10−71.5\times10^{-7} and 1.7×10−51.7\times10^{-5}. We do not yet have an explanation for this behavior.

In 3D, on uAu_A, the error with enrichment remains near 1.8×10−41.8\times10^{-4} across 4–32 branches, widths of 4–16, and one to four terms per edge. This limitation is consistent with the constant-amplitude edge terms, which restrict variation along the edge, as discussed in Section 6.2. Without enrichment, the best seed gives an error between 8.7×10−48.7\times10^{-4} and 1.3×10−31.3\times10^{-3} at every model size. However, at most sizes, one of the three seeds finishes with an error of 1.9×10−21.9\times10^{-2} or above.

Thanks for the authors' responses, I will keep my initial score.

Reviewer_fTwAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes enriched holomorphic neural representations for Laplace boundary-value problems with corner, edge, and boundary-condition singularities. In 2D, an entire complex-valued neural network is combined with singular functions selected from a geometry- and boundary-condition-dependent dictionary. Their real parts give a harmonic approximation trained using boundary data alone. The construction includes branch-cut prescriptions and optionally trainable singular exponents. In 3D, the model sums independent holomorphic networks evaluated along complex null directions and adds straight-edge or conical enrichment functions.

Validation includes closed-form singular benchmarks, the Motz and cracked-beam problems, a sampled suite of boundary configurations, and numerical CdZnTe detector examples. Matched 2D comparisons show substantial improvements from enrichment, including cases requiring a nontrivial neural remainder. The proposed model also achieves the lowest reported errors among the tested 3D methods, although these comparisons do not isolate the contribution of enrichment from changes to the harmonic network architecture.

Critical strengths
  1. A coherent representation that preserves harmonicity while incorporating singular behavior.

The construction combines an exactly harmonic neural representation with explicit singular functions and specifies how their branches are chosen for the tested geometries. The 3D null-direction construction also preserves the Laplace equation by design. This provides a clear mechanism for fitting boundary data without an interior PDE residual, while allocating explicit degrees of freedom to features that a smooth network may represent inefficiently. The treatment of multiple singular sites and the implementation details make the approach useful beyond a single illustrative corner problem.

  1. Strong matched evidence for the benefit of enrichment in 2D.

The plain and enriched holomorphic models use the same network and training budget, making their comparison particularly informative. On Motz’s problem and the cracked beam, the relative errors decrease from 5.4e-4 to 2.5e-8 and from 6.4e-4 to 8.6e-8, respectively (Table 2). Appendix C also reports that fitting the same four singular terms without the network gives errors of 5.5e-3 and 2.2e-2. These results support a useful interaction between explicit singular terms and the learned remainder. The sampled configurations and detector examples provide additional evidence that the approach can accommodate multiple singular sites.

Critical weaknesses
  1. The incremental benefit over closely related methods is insufficiently established, particularly in 3D.

The paper cites SEPINN, which already incorporates corner, mixed-boundary, and 3D edge singularities, but does not compare against it experimentally. Existing enriched holomorphic networks are also directly relevant to the contribution. Moreover, the discussion of trainable exponents should account for N’Guessan and Kim, “Discovering Scaling Exponents with Physics-Informed Müntz-Szász Networks” (ICML 2026), which studies learned Laplace corner exponents using harmonic power-law families.

In 3D, the proposed independent-branch architecture differs from the shared-network construction of the Ballini baseline, and there is no ablation retaining the proposed architecture while removing only the singular terms. This matters because, on the smooth cube where no enrichment is used, the proposed model already obtains 4.6e-6 versus 2.0e-4 for Ballini (Table 3), a roughly 43-fold difference. Consequently, the singular-problem comparisons cannot establish how much of the improvement comes from enrichment. A matched 3D ablation and a comparison against an existing singularity-enriched method are important for assessing the paper’s contribution.

  1. The dictionary in Table 1 does not cover some configurations allowed by the problem formulation.

Table 1 states that each junction matches one row. For a Neumann–Neumann corner, it assigns a power-law term when the fluxes are equal and the angle is re-entrant, or a logarithmic term for a flux jump when the angle belongs to πN. Other configurations receive no enrichment and are labeled as having bounded gradients.

Consider the sector 0 < r < 1, 0 < θ < 3π/2, with

u(r, θ) = r cos(θ) + r^(2/3) cos(2θ/3).

This function is harmonic. Its outward Neumann data on the two radial sides are 0 and 1, and its exact values can be prescribed on the circular arc as Dirichlet data. The fluxes are unequal and 3π/2 does not belong to πN, so the stated rule selects no enrichment. Nevertheless, the gradient magnitude of the second term is (2/3)r^(-1/3), which diverges at the corner.

This demonstrates that an inhomogeneous boundary-data contribution and a homogeneous corner singularity can coexist. The dictionary should explicitly incorporate a boundary-data lifting step and the remaining homogeneous singular modes, allow multiple enrichment families at a junction, or restrict its claimed scope. As written, this omission affects both the automatic selection rule and the premise that the network is left with a smooth remainder.

Questions for authors
  1. Attribution of the 3D improvement — additional evidence.

For a 3D singular problem with a nontrivial remainder, such as u_A or u_B, what is the error of exactly the same independent-branch model with Q = 0, using the same precision, samples, optimization budget, and restart selection? A corresponding comparison on the detector would also be informative. How does the enriched model compare with SEPINN when both methods receive comparable singularity information? These results could materially change my assessment of the incremental contribution.

  1. Coverage of the enrichment dictionary — clarification/correction.

How should Table 1 handle the unequal-flux re-entrant Neumann example described above? Is a lifting of inhomogeneous boundary data intended before selecting the corner modes? Please state the resulting selection rule, including whether a junction can receive both a boundary-jump term and a homogeneous fractional-power mode. This would clarify whether the issue is an omitted methodological step or a restriction on the supported problems.

  1. Detector evaluation — clarification.

Please reconcile the CZT metric definitions and values across Section 4.2, Table 2, and Table 6. In particular, are the Table 2 values relative L2 errors obtained by dividing the Table 6 RMSE values by the RMS of the reference solution, and are the evaluation points identical? Please also provide the grid or mesh settings and convergence evidence for the finite-difference and finite-element references, including the norm associated with the stated approximately 1e-4 reference error in 3D. These details are consequential because the detector improvements over the next-best method are comparatively modest.

1 comment

We thank the reviewer for their detailed feedback, particularly the counterexample to Table 1, which helped us correct the paper.

Weakness 1 and Question 1. Table R2 (general response) adds Ours, Q=0, removing only enrichment, with matched architecture, precision, samples, budget, restart selection and initial weights. Errors increase from 1.8×10−41.8\times10^{-4} to 1.5×10−21.5\times10^{-2} on uAu_A, and from 1.6×10−41.6\times10^{-4} to 1.2×10−21.2\times10^{-2} on uBu_B. On the detector, they increase from 1.3×10−21.3\times10^{-2} to 4.7×10−14.7\times10^{-1}. Across 20 seeds each for uA,uBu_A,u_B, enrichment wins on 39/40; even when the unenriched model fits the boundary data, error falls 4.3–14.5 times. On the smooth cube, the 43-fold gain over Ballini et al. cannot be attributed to enrichment. We will clarify this.

For SEPINN, we supplied the same singular sites, exponents and edge frames. We adapted its two-prism code, reproducing Examples 5.1–5.3 within a factor of 2. Our model is more accurate on every shared test: 12- and 14-fold on uA,uBu_A,u_B, 5.5-fold on the detector (Table R2), and on all seven 2D tests (Table R1). The MLP-core ablation retains terms, exponents and optimization, but uses a tanh MLP plus Laplace-residual loss. It is 4–5 times less accurate on uA,uBu_A,u_B and extruded Motz, and 2.3 times less accurate on the detector. It does better on the L-prism and conical notch, where terms represent the full solution.

Calafà et al. (2025) formulate their enriched holomorphic network for elasticity. With fixed exponents, our 2D model is close to theirs for Laplace's equation. Ours additionally supplies a geometry/BC-based dictionary, a branch-cut rule for multiple sites, and 3D edge/cone terms. We will cite N'Guessan and Kim (2026), who also train harmonic-power exponents; ours combines dictionary-selected terms with a harmonic remainder. Table R3 shows gains from perturbed initial exponents, not general identification of unknown exponents.

Weakness 2 and Question 2. The corrected rule selects homogeneous corner modes from angle and boundary-condition types, independently of data terms; both may occur at one junction. For straight-sided junctions with locally constant data, data terms and a linear particular solution handle the inhomogeneous part. Here g1,g2g_1,g_2 are Dirichlet values/outward fluxes at θ=0,ω\theta=0,\omega. Rule:

SidesHomogeneous exponents λk\lambda_kData term
D–Dkπ/ωk\pi/\omegaarg⁡w\arg w if g1≠g2g_1\ne g_2
N–Nkπ/ωk\pi/\omegawLog⁡ww\operatorname{Log}w if ω=mπ\omega=m\pi and g2≠(−1)m+1g1g_2\ne(-1)^{m+1}g_1
D–N(2k−1)π/(2ω)(2k-1)\pi/(2\omega)wLog⁡ww\operatorname{Log}w if ω=(2m−1)π/2\omega=(2m-1)\pi/2 and gN≠0g_N\ne0

Here m=1,2,…m=1,2,\ldots, k=1,…,Kk=1,\ldots,K; modes are added if λ1∉N\lambda_1\notin\mathbb N. Components/phases follow side orientation.

Table R5: changed instances (squares; boundary loss, three-seed medians). Gain = original-rule rerun / corrected loss.

InstanceSitesPlainCorrectedGain
Mixed D–N 034.6e-46.4e-817
Mixed D–N 148.6e-45.9e-75.6
Mixed D–N 222.2e-35.5e-79.9
Mixed D–N 328.3e-43.4e-75.2
Many segments 114.2e-34.4e-85.2
Many segments 228.8e-31.1e-50.8
Many segments 326.3e-31.6e-65.0

At 18 corners with nonzero data on both sides, selected terms plus a linear function match data below 10−1010^{-10}. At the reviewer's corner, it selects w2/3,w4/3w^{2/3},w^{4/3}, reducing error from 1.8×10−11.8\times10^{-1} to 9.2×10−109.2\times10^{-10}; the leading exponent is within 10−1010^{-10} of 2/32/3.

Only 7/24 instances change; six improve, up to 17-fold, and all seven remain orders below the unenriched model. We will update Table 1/Figure 5, state the scope (straight-sided 2D junctions with locally constant data), and clarify that finite enrichment need not leave a smooth remainder.

Question 3. Table 2 is Table 6 RMSE divided by reference-solution RMS (0.62 in units of 350 V). Both use the same runs and every-fourth interior reference-grid node. We will clarify this in Section 4.2.

The 2D reference uses five-point finite differences at h=0.025,0.0125h=0.025,0.0125 mm and first-order Richardson extrapolation. Refinements h=0.1/2kh=0.1/2^k mm (k=0,…,4k=0,\ldots,4) give orders 0.76, 0.91, 0.96. The relative L2L^2 difference to the next-finer combination (h=0.0125,0.00625h=0.0125,0.00625 mm), on the 0.10.1 mm grid's interior nodes, is 2.5×10−52.5\times10^{-5} (estimate, not bound).

The 3D reference uses graded Q1 FEM on the quarter box. Levels n=3,4,6,7 have 11.2,20.2,49.2,70.4 million unknowns; n counts graded cells per 0.1 mm beside each switch line. Successive relative L2 differences on 2M uniform points are 2.15e-4,1.58e-4,4.05e-5. Fitting en=Cn−pe_n=Cn^{-p} through differences on triples (3,4,6)/(4,6,7) gives p=1.91/1.37 and level-7 relative L2 error 1.0e-4–1.7e-4. These estimates are not bounds, but are over 60 times smaller than the detector performance gaps.

Reviewer_9rP1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper enriches holomorphic neural networks with singular functions chosen from local geometry and boundary conditions. The representation satisfies Laplace's equation by construction and is fitted using boundary data. Experiments cover 2D and 3D benchmarks and a radiation-detector application.

Critical strengths

The method gives substantial gains on Motz's problem and the cracked beam: errors decrease from 5.4e-4 and 6.4e-4 to 2.5e-8 and 8.6e-8 (Table 2). These results go beyond recovering a supplied exact solution and show that the network and singular terms can complement each other.

Critical weaknesses
  1. The additional contribution beyond known singular enrichment needs stronger evidence. Related neural constructions already exist, including SEPINN (Hu et al., 2024). On the four closed-form 2D benchmarks, the exact solution is already a supplied enrichment term. These cases demonstrate mode recovery. The more distinctive claim of learning unknown exponents is weakly tested: initialization usually uses theoretical values, and training exponents changes the median boundary-loss ratio only to 0.98 in Table 7. I would like clearer evidence of what the method adds when the singular structure is not already accurately specified.

  2. The source and practical value of the 3D gains are unclear. There is no matched ablation removing enrichment from the authors' own 16-branch model. On the smooth cube, without enrichment, that model already improves on the earlier holomorphic baseline by about 43 times (Table 3). The other 3D comparisons therefore do not isolate enrichment. Comparisons with classical solvers using similar singularity knowledge are also missing, leaving the practical benefit difficult to judge.

Questions for authors
  1. How does the same 3D network perform without enrichment, under the same training budget, on the extruded Motz problem and detector?

  2. Do trainable exponents remain useful when initialized substantially away from their theoretical values?

  3. On a representative 2D problem, how does accuracy versus total runtime compare with a singular-series method or a suitable classical Laplace solver?

2 comments

We thank the reviewer for their feedback and we aim to address all the weaknesses and questions highlighted by the reviewer.

Question 1: We added ablations to Table R2 (general response), removing either the enrichment terms (Ours, Q=0Q=0) or the holomorphic core (MLP core). On Extruded Motz, the error increases from 3.4×10−63.4\times10^{-6} to 8.6×10−48.6\times10^{-4} and 1.5×10−51.5\times10^{-5}, respectively, confirming the contribution of both components.

Question 2: Table R3 (general response) tests exponent training by initializing all exponents at mm times their theoretical values, then either fixing or training them. Training improves accuracy for all perturbed initializations, demonstrating its usefulness.

Question 3: We measured runtime and accuracy. Classical solvers are faster, and the singular series is more accurate. On one 24-core AMD EPYC 9654 allocation, Motz results are: singular series (Li's boundary approximation method, 40 terms), 5.4×10−135.4\times10^{-13} in 0.08 s; Lightning, 6.5×10−76.5\times10^{-7} in 4.4 s; our model, 1.3×10−71.3\times10^{-7} in 15 minutes (budget 1) and 2.5×10−82.5\times10^{-8} in 26 minutes (budget 2). On the cracked beam: series, 1.1×10−121.1\times10^{-12} in 0.08 s; Lightning, 1.9×10−71.9\times10^{-7} in 4.5 s; our model, 3.1×10−73.1\times10^{-7} in 16 minutes and 2.1×10−72.1\times10^{-7} in 26 minutes. Neural timings include JIT compilation, three restarts and selection, but exclude scoring and budget selection. Budget 3 was not timed; the 8.6×10−88.6\times10^{-8} result in Table R1 comes from this budget.

We therefore do not position our method as a replacement for established numerical solvers. Instead, its mesh-free, differentiable representation offers different properties, potentially useful when harmonic solutions are components of larger differentiable models or continuous evaluation and spatial derivatives are required.

Weakness 1: We added component ablations and further experiments (Tables R1 and R2). While closed-form 2D problems mainly test recovery of supplied enrichment terms, on Motz and the cracked beam the terms alone give errors of 5.5×10−35.5\times10^{-3} and 2.2×10−22.2\times10^{-2}, compared with 2.5×10−82.5\times10^{-8} and 8.6×10−88.6\times10^{-8} when combined with the holomorphic network (Table 5).

We also added four variants with smooth harmonic components outside the enrichment span: L-shape and pure switch add ℜez\Re e^z and ℜ(z−1/2)1/2\Re(z-1/2)^{1/2}, respectively; the 3π/23\pi/2 sector and Motz add ℜ[1/(z−zp)]\Re[1/(z-z_p)] and ℜz1/2\Re z^{1/2}, respectively, with zp=1.1eiπ/4z_p=1.1e^{i\pi/4} and zp=0.3−0.25iz_p=0.3-0.25i. All additions have unit coefficient; the Motz variant uses only the leading mode. Terms alone give errors between 1.4×10−11.4\times10^{-1} and 9.3×10−19.3\times10^{-1}, while our model achieves 1.4×10−81.4\times10^{-8} to 5.4×10−65.4\times10^{-6}.

SEPINN receives the same singular information but is less accurate on all seven tested 2D problems (Table R1); on Motz, it gives 3.8×10−23.8\times10^{-2} versus 2.9×10−82.9\times10^{-8} for our model. Finally, the exponent ratio of 0.98 in Table 7 uses theoretical initialization and therefore does not test recovery from perturbed values. Table R3 addresses this through the experiment described in Question 2.

Weakness 2: We added the matched ablation Ours, Q=0Q=0 in Table R2, using the same 16-branch model, seeds, samples, budget and restart rule, with only enrichment removed. On the smooth cube, which has no singular sites, separate GPU executions use matched samples and initial core weights. The small difference (medians 3.9×10−63.9\times10^{-6} and 4.6×10−64.6\times10^{-6}; per-seed ranges 3.73.7--4.5×10−64.5\times10^{-6} and 4.44.4--4.9×10−64.9\times10^{-6}) is due to floating-point nondeterminism. Thus, the factor of 43 over Ballini et al. cannot be attributed to enrichment, which we will clarify.

On uAu_A, uBu_B, Extruded Motz and the detector, removing enrichment increases error by approximately 36--250 times. For example, errors increase from 1.8×10−41.8\times10^{-4} to 1.5×10−21.5\times10^{-2} on uAu_A, and from 1.3×10−21.3\times10^{-2} to 4.7×10−14.7\times10^{-1} on the detector. Across 20 seeds each for uAu_A and uBu_B, the enriched model performs better on 39 of 40 seeds. We address classical solver comparisons in Question 3.

Thank you for the additional experiments, which address my main technical concerns. However, I remain unconvinced that the advance over existing singularity-enrichment approaches is substantial enough to warrant acceptance. I therefore maintain my rating of 2.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Addressing questions and feedback raised by multiple reviewers

We thank the reviewers for their time and valuable feedback. Since several reviewers raised similar points, we provide this general response for all reviewers.

We ran additional experiments to assess our method against state-of-the-art methods and classical solutions. The results below will be included in the camera-ready version.

Table R1. 2D relative L2L^2 error (relative D0D_0 error where marked); median of three seeds. Plain: model without terms. Terms only: least-squares fit of singular terms, without a network. Published solvers use 5,000 shared evaluation points at their best schedule; PIHNN and Lightning use the authors' code. Ablation and Ours results retain the original evaluations, with the best budget selected separately for each metric. Shared-point Ours results are in footnote aa.

ProblemPlainTermsPINNSEPINNHNNPIHNNLightningOurs$^a$
Sector 3π/23\pi/27.7e-45.5e-165.6e-35.9e-31.0e-21.6e-25.7e-123.3e-12
Sector 7π/47\pi/41.9e-29.9e-161.5e-28.9e-31.4e-11.3e-15.1e-71.4e-16
L-shape7.5e-45.3e-164.4e-34.9e-31.1e-22.2e-22.9e-122.9e-16
Slit disk3.2e-12.0e-162.0e-11.5e-13.5e-13.9e-15.92.8e-12
Pure switch mode7.6e-32.0e-151.8e-22.3e-23.4e-22.8e-21.4e-71.1e-11
Motz5.4e-45.5e-33.7e-23.8e-22.7e-23.8e-26.5e-72.5e-8
Cracked beam6.4e-42.2e-21.2e-21.6e-22.6e-23.2e-21.9e-78.6e-8
Motz, D0D_04.2e-44.4e-31.6e-23.1e-26.3e-29.5e-21.6e-68.2e-7
Cracked beam, D0D_05.9e-41.1e-21.3e-29.1e-36.2e-27.6e-24.8e-72.0e-6

a^a On shared points, Ours gives 3.4e-12, 2.4e-16, 3.2e-16, 3.0e-12 and 1.1e-11 on the closed-form problems, and 2.9e-8 and 1.8e-7 on Motz and the cracked beam. At budgets selected by solution L2L^2 error (2 and 3), their D0D_0 errors are 2.3e-6 and 2.7e-6; the main column reports the best D0D_0 errors at budgets 3 and 1. Coefficient errors depend on budget, not evaluation points.

Table R2. 3D relative L2L^2 error; median of three seeds. Ours, Q=0Q=0: terms removed, with all other settings identical to Ours. MLP core: holomorphic core replaced by a tanh MLP with Laplace-residual loss; best of four settings (width 64/32, learning rate 1e-2/1e-3). At width 32, it has as many parameters as our core. SEPINN uses the same singular sites, theoretical exponents and edge frames. Baselines are as in Table 3, float32 at their best schedule.

ProblemBest baselineSEPINNOurs, Q=0Q=0MLP coreOurs
L-prism$^b$8.7e-41.2e-32.0e-22.7e-126.9e-9
uAu_A8.9e-42.1e-31.5e-27.2e-41.8e-4
uBu_B1.1e-32.3e-31.2e-27.1e-41.6e-4
Smooth cube1.5e-5n/a3.9e-62.2e-54.6e-6
Conical notch$^b$1.2e-3n/a6.0e-2$^a$2.4e-126.0e-9
Extruded Motz4.1e-31.6e-28.6e-41.5e-53.4e-6
CZT detector2.4e-27.2e-24.7e-13.0e-21.3e-2

a^a At learning rate 1e-3; at 1e-2 and 3e-3, the model without terms does not fit the boundary data. b^b The terms represent the whole solution, so the core only needs to stay near zero.

Unless stated otherwise, experiments use the submitted paper's settings. In 2D, restart selection uses the lowest boundary loss on a fresh sample four times denser than training; non-finite restarts are excluded.

On Motz and the cracked beam, combining the holomorphic network and enrichment improves over either component alone. The matched 3D ablation also shows the benefit of enrichment when a nontrivial remainder is required. Our method also outperforms SEPINN in all tested examples. We did not evaluate SEPINN on the smooth cube (no singularity) or conical notch (our implementation has no corresponding singular function).

Furthermore, training the exponents improves accuracy even when they are initialized far from their theoretical values.

Table R3. Relative L2L^2 solution error when exponents start away from theory. On Motz and the cracked beam, all four exponents start at mm times their theoretical values and are either fixed (pinned) or trained. Results are at budget 1 (Table 2 reports the best budget); medians of three seeds, each with three restarts.

mm0.50.750.91.01.11.251.5
Motz, pinned7.3e-51.6e-44.7e-51.3e-78.6e-53.1e-44.2e-4
Motz, trained1.3e-63.0e-72.7e-72.6e-73.0e-74.3e-79.6e-7
Cracked, pinned4.8e-52.1e-47.4e-53.3e-78.7e-53.1e-45.1e-4
Cracked, trained6.1e-71.8e-71.3e-71.2e-71.6e-71.2e-76.9e-7

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.