Unverified 2026

Gap-Graded Neural Collocation

Implementation & benchmark of arXiv:2607.22128 — High-order finite element method for perfect conductivity and linear elasticity with nearly touching inclusions

Usefulness5/10

Difficulty4/10

Novelty6/10

Training Accuracy Sample efficiency Stability Training Mlp Geometry Numerical analysis Pde Approximation theory Geometry

Source paper: High-order finite element method for perfect conductivity and linear elasticity with nearly touching inclusions arXiv:2607.22128 ⓘ · analyzed Aug 30, 2026

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 Not peer-reviewed.

Idea description

Replace uniform PINN or neural-operator collocation by a graded point distribution concentrated in narrow regions between nearly touching interfaces. Use the paper's distance-dependent mesh scale to determine point spacing, and switch to a gap-dependent minimum scale when the separation becomes too small for the global mesh.

Formulas

$$\Omega_*:=\{x\in\Omega_1:|x'|\leq\varepsilon^{1/2}\}.$$

$$\hbar(x)=\begin{cases}h,&x\in\Omega_0,\\ \varepsilon^{\alpha/2}h,&x\in\Omega_*,\\ |x'|^\alpha h,&\text{otherwise}.\end{cases}$$

$$\varepsilon\geq(\kappa h)^{\frac{1}{1-\alpha/2}}\quad\text{or}\quad\varepsilon<(\kappa h)^{\frac{1}{1-\alpha/2}}.$$

$$\mathcal L_{\mathrm{weighted}}(\theta)=\frac{1}{N}\sum_{i=1}^N\frac{r_\theta(x_i)^2}{q(x_i)},\qquad q(x)\propto \hbar(x)^{-n},$$

Mathematical statement

The paper prescribes the local mesh size $\hbar(x)$ by equation (3.1): $\hbar(x)=h$ in the regular region $\Omega_0$, $\hbar(x)=\varepsilon^{\alpha/2}h$ in the narrow core $\Omega_*$, and $\hbar(x)=|x'|^\alpha h$ elsewhere. Here $h$ is the coarse mesh size, $\varepsilon>0$ is the minimum separation between the two inclusions, $x'$ is the tangential coordinate measuring distance from the centerline of the gap, and $\alpha$ controls grading strength. The narrow core is $\Omega_* = \{x\in\Omega_1: |x'|\leq\varepsilon^{1/2}\}$, where $\Omega_1$ denotes the inter-inclusion region. The construction distinguishes $\varepsilon\geq(\kappa h)^{1/(1-\alpha/2)}$ from $\varepsilon<(\kappa h)^{1/(1-\alpha/2)}$, with $\kappa>0$ a fixed constant. We transfer the principle by assigning collocation density proportional to $\hbar(x)^{-n}$, where $n\in\{2,3\}$ is the spatial dimension, and importance-weighting the squared PDE residual by the inverse sampling density. The mathematical benefit is targeted resolution of the high-derivative region without refining the entire domain.

Implementation notes

(1) Exact integration point: modify the collocation sampler and interior residual loss of a PINN solving an elliptic conductivity or elasticity problem with two inclusions. The network remains an MLP mapping coordinates $x$ to potential, displacement, or stress; only point allocation and loss weighting change. Given the inclusion boundaries, estimate the minimum gap $\varepsilon$, identify the gap centerline, and compute the tangential coordinate $x'$ for each candidate point. (2) Pseudocode: choose coarse scale $h$, exponent $\alpha\in(0,1)$, constant $\kappa$, and batch size $N$. Compute $\varepsilon$. Define $\Omega_*$ by $|x'|\leq\sqrt{\varepsilon}$, evaluate $\hbar(x)$ with the paper's piecewise formula, and form an unnormalized density $\tilde q(x)=\hbar(x)^{-n}$. Estimate its normalization by region-wise Monte Carlo. Sample region labels with probabilities proportional to estimated region volume times mean $\tilde q$, then sample uniformly inside each region. For each point calculate the PDE residual $r_\theta(x)$ using automatic differentiation and add $r_\theta(x)^2/q(x)$ to the interior loss; retain ordinary boundary-condition losses. Resample every 100 to 500 optimizer steps. (3) Taken directly from the mathematics: the core radius $\sqrt{\varepsilon}$, the scale law, and the threshold $\varepsilon=(\kappa h)^{1/(1-\alpha/2)}$. Estimated empirically: region volumes, normalization of $q$, and the best $\alpha$, initially testing 0.5, 0.75, and 1.0. If geometry is supplied only as a point cloud, estimate the gap center and $\varepsilon$ using nearest boundary-point pairs. (4) First cheap experiment: train a tanh or sine PINN on a 2D conductivity problem with two circular inclusions at separations $10^{-1},10^{-2},10^{-3}$. Compare uniform sampling, residual-adaptive sampling, and gap-graded sampling with the same number of PDE evaluations and the same optimizer. Report relative $H^1$ error, maximum electric-field error in the gap, final residual, wall-clock time, and seed variance. Success is lower gap-field error at equal budget, or equal error with at least half as many collocation points.

Verification

This idea has not been verified yet.

Verification happens in two stages: Stage 1 — a mechanism check on a toy system confirms the claimed mathematical phenomenon reproduces; Stage 2 — a benchmark implements the idea on a real (small) neural network task and compares it against a tuned baseline over 8 paired seeds with a permutation test.

Artifacts

Artifacts unavailable.