# Exact-boundary geometric PINN ansatz

- ID: 1550
- Canonical URL: https://synthcore.org/idea/1550/exact-boundary-geometric-pinn-ansatz
- API JSON: https://synthcore.org/api/idea/1550.json
- API Markdown: https://synthcore.org/api/idea/1550.md
- Verification status: mech_ok_no_baseline
- Source: [arXiv:2607.28733](https://arxiv.org/abs/2607.28733)
- Category: architecture
- Solves: stability, accuracy, sample-efficiency
- ML areas: architecture, loss, training
- Math tags: differential-geometry, geometry, pde
- Ratings: usefulness 7/10; difficulty 4/10; novelty 6/10

## Idea description

Replace penalty enforcement of boundary and asymptotic conditions by a neural-network parametrization that satisfies them identically. For a disc mapped to hyperbolic half-space, construct the positive height coordinate with a vanishing boundary factor and construct tangential coordinates as a fixed boundary extension plus a vanishing correction.

## Mathematical statement

The geometric problem uses the hyperbolic half-space metric $g_{\mathrm{hyp}}=X^{-2}(dX^{2}+|d\boldsymbol{Y}|^{2})$, whose ideal boundary is $X=0$. The unknown map is $u=(X,\boldsymbol{Y}):D^{2}\to\overline{\mathbb H}^{n+1}$ with $u|_{\partial D^{2}}=\gamma$, where $\gamma:S^{1}\to\mathbb R^{n}$ is prescribed and the interior should satisfy the vanishing tension equation $\tau(u)=\operatorname{tr}_{u^{*}g_{\mathrm{hyp}}}\nabla du=0$. Let $r=|z|$ be the disc radius, let $\rho(r)=1-r^{2}$, let $E[\gamma](z)$ be any fixed extension of the boundary curve into the disc satisfying $E[\gamma](e^{i\theta})=\gamma(e^{i\theta})$, and let $a_{\theta}(z)$ and $b_{\theta}(z)$ be unconstrained neural outputs. The adapted ansatz $X_{\theta}(z)=\rho(r)\exp(a_{\theta}(z))$ is strictly positive in the interior and exactly zero on the boundary, while $\boldsymbol{Y}_{\theta}(z)=E[\gamma](z)+\rho(r)b_{\theta}(z)$ exactly equals $\gamma$ on $\partial D^{2}$. The PDE residual can then use the pullback metric $g_{ab}=X^{-2}\sum_{\alpha}\partial_{a}u^{\alpha}\partial_{b}u^{\alpha}=X^{-2}(J^{\top}J)_{ab}$ and $R_{\theta}=\tau(u_{\theta})$ without boundary penalty terms.

## Key formulas

- $$g_{\mathrm{hyp}}=X^{-2}\left(dX^{2}+|d\boldsymbol{Y}|^{2}\right),\qquad \partial\overline{\mathbb H}^{n+1}=\{X=0\}\cong\mathbb R^{n}.$$
- $$u=(X,\boldsymbol{Y}):D^{2}\to\overline{\mathbb H}^{n+1},\qquad u|_{\partial D^{2}}=\gamma,\qquad \tau(u)=\operatorname{tr}_{u^{*}g_{\mathrm{hyp}}}\nabla du=0.$$
- $$X_{\theta}(z)=(1-|z|^{2})\exp(a_{\theta}(z)),\qquad \boldsymbol{Y}_{\theta}(z)=E[\gamma](z)+(1-|z|^{2})b_{\theta}(z).$$
- $$\mathcal L(\theta)=\frac{1}{N}\sum_{i=1}^{N}\left\|\tau\!\left(u_{\theta}(z_i)\right)\right\|_{2}^{2},\qquad g_{ab}(z)=X_{\theta}(z)^{-2}\sum_{\alpha}\partial_a u_{\theta}^{\alpha}(z)\partial_b u_{\theta}^{\alpha}(z).$$

## Implementation notes

Integrate this at the output head of a PINN whose inputs are disc coordinates $z=(z_1,z_2)$ and whose raw outputs are one scalar $a_\theta(z)$ and one vector $b_\theta(z)\in\mathbb R^n$. Precompute a differentiable boundary extension $E[\gamma]$; for a closed planar curve, a simple choice is the harmonic or Fourier extension obtained by evaluating the truncated Fourier series of $\gamma(e^{i\phi})$. At every collocation point compute $\rho=1-z_1^2-z_2^2$, then set $X=\rho\exp(a)$ and $Y=E[\gamma](z)+\rho b$. Compute first and second derivatives of $(X,Y)$ with respect to $z$ and form the pullback metric $g=X^{-2}J^\top J$; invert the $2\times2$ matrix $g$ with a stable determinant clamp. Assemble the hyperbolic tension residual using $\tau^\alpha=g^{ab}(\partial_{ab}u^\alpha+\Gamma^\alpha_{\beta\gamma}(u)\partial_a u^\beta\partial_b u^\gamma-\Gamma^c_{ab}(g)\partial_cu^\alpha)$, or use the coordinate expression already implemented for the target PDE. Optimize only the mean squared residual, with no boundary loss and no separate positivity penalty. Compare against an unconstrained network trained with boundary and positivity penalties on a small hyperbolic minimal-disc problem. Check boundary error, minimum predicted $X$, residual loss, and convergence variance across random seeds; success means boundary error remains at machine precision while residual reaches the same or lower value with fewer optimization steps.

## Verification

- Status: mech_ok_no_baseline
- Mechanism evidence: yes
- Mechanism confirmed: yes
- Practical verdict: no_effect
- Verdict: Built an exact-boundary PINN head with X=(1-|z|²)exp(a) and Y=E[gamma]+(1-|z|²)b, plus a matched penalty baseline and training experiment. The mechanism checks confirmed machine-precision boundary enforcement across raw-output scales, linear correction scaling with the vanishing factor (observed coefficient 3.25, predicted 3.25), and strict positivity of X. In the toy training problem, the exact ansatz reached residual MSE 4.5e-5–7.7e-5 versus 1.45e-2–1.72e-2 for the baseline, while eliminating boundary violations and negative X.

### Mechanism check

- Verdict: Built an exact-boundary PINN head with X=(1-|z|²)exp(a) and Y=E[gamma]+(1-|z|²)b, plus a matched penalty baseline and training experiment. The mechanism checks confirmed machine-precision boundary enforcement across raw-output scales, linear correction scaling with the vanishing factor (observed coefficient 3.25, predicted 3.25), and strict positivity of X. In the toy training problem, the exact ansatz reached residual MSE 4.5e-5–7.7e-5 versus 1.45e-2–1.72e-2 for the baseline, while eliminating boundary violations and negative X.
- Confidence: 8/10
- Limitations: The training comparison uses a small manufactured Laplacian residual proxy rather than the full hyperbolic harmonic-map tension equation and pullback Christoffel terms. It does not test complex target geometries, Fourier extensions beyond the simple identity boundary curve, larger networks, or equal-FLOP convergence curves; float32 training evaluation shows about 1e-7 boundary error due to coordinate arithmetic.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 0.7926
- Paired wins: 5/8
- Benchmark verdict: no significant win

## Artifacts

- [bench_exact_boundary.py](https://synthcore.org/code/474/bench_exact_boundary.py)
- [bench_report.json](https://synthcore.org/code/474/bench_report.json)
- [poisson_boundary_track.py](https://synthcore.org/code/474/poisson_boundary_track.py)
- [report.md](https://synthcore.org/code/474/report.md)
- [report_bench_2026-08-31T143044.md](https://synthcore.org/code/474/report_bench_2026-08-31T143044.md)
- [results.json](https://synthcore.org/code/474/results.json)
- [run_experiment.py](https://synthcore.org/code/474/run_experiment.py)
- [Download all files as ZIP](https://synthcore.org/download/474)

## Disclaimer

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