# Singularity-Enriched Neural Ansatz

- ID: 363
- Canonical URL: https://synthcore.org/idea/363/singularity-enriched-neural-ansatz
- API JSON: https://synthcore.org/api/idea/363.json
- API Markdown: https://synthcore.org/api/idea/363.md
- Verification status: beats_baseline
- Source: [arXiv:2607.02334](https://arxiv.org/abs/2607.02334)
- Category: architecture
- Solves: accuracy, sample-efficiency, scalability
- ML areas: loss, geometry, training, data-augmentation
- Math tags: pde, approximation-theory, numerical-analysis, functional-analysis
- Ratings: usefulness 6/10; difficulty 5/10; novelty 7/10

## Idea description

Add an explicit local power-law singular basis to a neural field near mixed Dirichlet-Neumann junctions, allowing the neural network to learn only the smoother remainder. Use the predicted or fitted singular exponent to concentrate collocation points near the junction. This directly targets the regularity bottleneck identified by the paper, where increasing polynomial degree or network capacity cannot overcome a convergence cap under uniform resolution.

## Mathematical statement

Near a mixed-boundary junction \(\bm{x}_0\), the paper gives the Kondratiev expansion \(\bm{u}(r,\theta)=\kappa r^{\lambda}\bm{\Phi}(\theta)+\bm{w}\), where \(r=\|\bm{x}-\bm{x}_0\|\), \(\theta\) is the polar angle, \(\kappa\) is an unknown amplitude, \(\bm{\Phi}(\theta)\) is an angular eigenfunction, and \(\bm{w}\) is locally smoother with \(\bm{w}\in(H^{1+\lambda+\varepsilon})^d\) for \(\varepsilon>0\). The exponent \(\lambda=\lambda(\omega,\nu)\in(0,1]\) is the smallest positive root of the Lamé characteristic equation, determined by the opening angle \(\omega\) and Poisson ratio \(\nu\); for a straight mixed junction \(\omega=\pi\), \(\lambda=1/2\). The paper's radial grading uses local element size \(h_T\simeq h^{1/\mu}\) for \(r_T\le h^{1/\mu}\) and \(h_T\simeq h r_T^{1-\mu}\) otherwise, where \(\mu\in(0,1]\), \(r_T\) is distance from the junction, and smaller \(\mu\) means stronger refinement. Adapt the representation to \(\bm{u}_{\theta}(x)=\chi(r)r^{\hat\lambda}(a_{\theta}\bm{\Phi}_{\psi}(\theta)+\bm{v}_{\theta}(x))+(1-\chi(r))\bm{v}_{\theta}^{\mathrm{far}}(x)\), where \(\hat\lambda\) is the chosen exponent, \(a_{\theta}\) is a learned amplitude, \(\bm{\Phi}_{\psi}\) is a learned or analytical angular mode, \(\bm{v}_{\theta}\) is a local neural remainder, and \(\chi\) is a smooth cutoff.

## Key formulas

- $$\bm{u}(r,\theta)=\kappa\,r^{\lambda}\,\bm{\Phi}(\theta)+\bm{w},\qquad\bm{w}\in\big(H^{1+\lambda+\varepsilon}\big)^{d}\text{ locally},\ \varepsilon>0$$
- $$\lambda=\lambda(\omega,\nu)\in(0,1]$$
- $$h_{T}\ \simeq\ \begin{cases}h^{1/\mu},&r_{T}\le h^{1/\mu},\\ h\,r_{T}^{\,1-\mu},&r_{T}>h^{1/\mu},\end{cases}$$
- $$\bm{u}_{\theta}(x)=\chi(r)\,r^{\hat\lambda}\big(a_{\theta}\bm{\Phi}_{\psi}(\theta)+\bm{v}_{\theta}(x)\big)+(1-\chi(r))\bm{v}_{\theta}^{\mathrm{far}}(x)$$

## Implementation notes

(1) Exact integration point: modify the output head of a PINN or neural operator for a domain with known mixed-boundary corners. For each junction \(\bm{x}_0\), compute differentiable polar coordinates \(r=(\|x-\bm{x}_0\|^2+\epsilon_r^2)^{1/2}\) and \(\theta=\operatorname{atan2}(x_2-x_{0,2},x_1-x_{0,1})\). Use a smooth cutoff \(\chi(r)\) that is one for \(r<r_0/2\) and zero for \(r>r_0\). (2) Pseudocode: choose \(\hat\lambda\) from the analytical corner problem when available; compute `sing=r**lambda*(a*Phi(theta)+v_local(x))`; compute `u=chi*sing+(1-chi)*v_far(x)`; evaluate the governing energy or PDE residual and train all network parameters, amplitude parameters, and angular coefficients. Represent \(\bm{\Phi}_{\psi}\) with a small angular MLP or Fourier features, while enforcing the appropriate angular boundary behavior if known. (3) The mathematical input is the power-law singularity and radial grading law. Estimate \(\hat\lambda\) empirically when the characteristic equation is unavailable by fitting the slope of \(\log |u-u_{\mathrm{smooth}}|\) versus \(\log r\) during warm-up, then clip it to \([0.1,1.0]\). Draw more collocation points from annuli near the corner according to the grading profile, while retaining a uniform far-field batch. (4) First cheap experiment: solve a manufactured Poisson or elasticity problem on a square with one Dirichlet-Neumann junction. Compare a standard MLP PINN, a uniformly oversampled PINN, and the enriched-plus-graded model at equal point evaluations. Track relative \(L^2\) error, corner error, residual variance, and error versus sample count. Success is at least 2x lower global error at equal samples and a visibly improved convergence slope near the junction.

## Verification

- Status: beats_baseline
- Mechanism evidence: yes
- Mechanism confirmed: yes
- Practical verdict: helps
- Verdict: Built a differentiable singularity-enriched neural ansatz with a smooth cutoff, learned amplitude, fixed lambda=1/2 mode, and corner-focused sampling. The numerical math check recovered log-log slope 0.500 exactly, while grading increased the fraction of points within r<0.12 from 1.12% to 8.79%. At equal 1,600 training points, enriched+graded reduced global relative L2 error from 1.475% for the uniform baseline to 0.826%, and reduced corner relative error from 21.5% to 2.90% (about 7.4x better), so the local regularity-bottleneck effect is real in this toy setting.

### Mechanism check

- Verdict: Built a differentiable singularity-enriched neural ansatz with a smooth cutoff, learned amplitude, fixed lambda=1/2 mode, and corner-focused sampling. The numerical math check recovered log-log slope 0.500 exactly, while grading increased the fraction of points within r<0.12 from 1.12% to 8.79%. At equal 1,600 training points, enriched+graded reduced global relative L2 error from 1.475% for the uniform baseline to 0.826%, and reduced corner relative error from 21.5% to 2.90% (about 7.4x better), so the local regularity-bottleneck effect is real in this toy setting.
- Confidence: 8/10
- Limitations: This is a manufactured scalar regression rather than a full mixed-boundary elasticity or Poisson PINN with PDE residuals. The angular mode and exponent were supplied analytically, only one fixed seed and one sample budget were tested, and no wall-clock or FLOP-normalized comparison was recorded. The global improvement was less than the suggested 2x threshold, although the corner improvement was substantially larger.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 0.0081
- Paired wins: 8/8
- Benchmark verdict: idea better (significant)

## Artifacts

- [bench_report.json](https://synthcore.org/code/1276/bench_report.json)
- [check_signature.py](https://synthcore.org/code/1276/check_signature.py)
- [experiment.py](https://synthcore.org/code/1276/experiment.py)
- [mixed_junction_track.py](https://synthcore.org/code/1276/mixed_junction_track.py)
- [report.md](https://synthcore.org/code/1276/report.md)
- [report_bench_2026-09-04T003530.md](https://synthcore.org/code/1276/report_bench_2026-09-04T003530.md)
- [run_bench.py](https://synthcore.org/code/1276/run_bench.py)
- [Download all files as ZIP](https://synthcore.org/download/1276)

## Disclaimer

AI-generated research hypothesis, automatically tested. Not peer-reviewed.