✓✓ Beats tuned baseline 2026

Energy-trained monotone coordinate warp

Implementation & benchmark of arXiv:2607.23229 — Mechanics-trained neural coordinate mapping for B-spline analysis of crack-tip and corner singularities

Usefulness7/10

Difficulty5/10

Novelty7/10

Geometry Accuracy Stability Sample efficiency Mlp Training Loss Geometry Numerical analysis Optimization Differential geometry

Source paper: Mechanics-trained neural coordinate mapping for B-spline analysis of crack-tip and corner singularities arXiv:2607.23229 ⓘ · analyzed Aug 30, 2026

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 Not peer-reviewed.

Idea description

Replace raw spatial coordinates supplied to a neural field or PINN by a learnable monotone radial coordinate generated from a positive neural density. The density is trained through the PDE energy or residual after solving for the network weights, allowing the warp to discover where resolution is needed without singularity labels or an analytic interior solution. Near a singular point, a factor s^(q-1) gives a controlled regularity gain, while a positive learned correction redistributes resolution away from the point.

Formulas

$$r_{\phi}(s)=\frac{\int_{0}^{s}\rho_{\phi}(\xi)\,\mathrm{d}\xi}{\int_{0}^{1}\rho_{\phi}(\xi)\,\mathrm{d}\xi},\qquad r_{\phi}^{\prime}(s)=\frac{\rho_{\phi}(s)}{\int_{0}^{1}\rho_{\phi}(\xi)\,\mathrm{d}\xi}>0.$$

$$\rho_{\phi}(s)\sim C_{\phi}s^{q-1}\ \Longrightarrow\ r_{\phi}(s)\sim\widetilde{C}_{\phi}s^{q},\qquad r^{\lambda}\mapsto \widetilde{C}_{\phi}^{\lambda}s^{q\lambda}.$$

$$\boldsymbol{\chi}_{\phi}(s,a)=r_{\phi}(s)(\cos a,\sin a),\qquad J_{\phi}(s)=r_{\phi}(s)r_{\phi}'(s).$$

$$\mathcal{E}(\theta,\phi)=\int_{\widehat{\Omega}}\frac{1}{2}\left\|\nabla_{\phi}u_{\theta}(s,a)\right\|^{2}J_{\phi}(s)\,\mathrm{d}s\,\mathrm{d}a-\int_{\widehat{\Omega}}f(\boldsymbol{\chi}_{\phi}(s,a))u_{\theta}(s,a)J_{\phi}(s)\,\mathrm{d}s\,\mathrm{d}a.$$

Mathematical statement

The paper defines the computational radius s in [0,1], an unnormalized positive density rho_phi(s)>0, and the normalized map r_phi(s) = integral from 0 to s of rho_phi(xi) dxi divided by the integral from 0 to 1 of rho_phi(xi) dxi, where phi are density parameters. Its derivative is r_phi'(s) = rho_phi(s) divided by the integral from 0 to 1 of rho_phi(xi) dxi, so r_phi(0)=0, r_phi(1)=1, and the map cannot fold. The density is parameterized to satisfy rho_phi(s) approximately C_phi s^(q-1) near the singular point, with q >= 1, implying r_phi(s) approximately C_tilde_phi s^q. Therefore a physical singular term r^lambda, where lambda > 0 is the singular exponent, becomes C_tilde_phi^lambda s^(q lambda), increasing computational regularity when q lambda > lambda. In the paper's polar chart chi_phi(s,a) = r_phi(s)(cos(a), sin(a)), the Jacobian is J_phi = r_phi r_phi', and scalar derivatives transform as grad_phi v = (v_s / r_phi') e_r + (v_a / r_phi) e_a. The machine-learning adaptation trains phi by differentiating an outer PDE objective through inner neural-field optimization: phi is not learned from an unknown exponent label, but from the discrete energy attained by the network.

Implementation notes

Integrate the warp immediately before the neural-field input. For a two-dimensional singularity at x_0, sample computational points (s,a), map them to physical points x = x_0 + r_phi(s)(cos(a), sin(a)), and feed either (s,a) or the mapped Cartesian coordinates to an MLP u_theta. Use a positive density with a hard singular factor: rho_phi(s) = epsilon + s^(q-1) exp(h_phi(s)), where h_phi is a small MLP, q >= 1 is fixed initially, and epsilon > 0 prevents numerical zero density away from the collapsed point. Approximate cumulative integrals on a fixed quadrature grid and compute r_phi with a cumulative sum, followed by interpolation at random samples. Pseudocode: initialize theta and phi; compute rho_i = epsilon + s_i^(q-1) exp(h_phi(s_i)); compute cumulative sums R_i = sum over j <= i of rho_j times delta_s; set r_i = R_i / R_N; map every sample to x_i; evaluate u_theta; compute transformed derivatives using u_s / r_phi' and u_a / r_phi; form the weighted energy using J = r_phi r_phi'; update theta for several inner steps; then update phi using the energy or a held-out residual after unrolling one to five inner steps. Clip h_phi to a bounded range and add only a weak smoothness penalty on h_phi', not a positivity penalty. The quantities computed directly from the mathematics are the normalized monotone map, its derivative, the Jacobian, and the regularity exponent q lambda; the useful redistribution is evaluated empirically through the outer energy. First test a small Deep Ritz MLP on Poisson's equation over an L-shaped domain with a re-entrant-corner singularity. Compare identity coordinates, a fixed r = s^2 warp, and the learned correction at equal MLP size and sample count. Measure energy-norm error against a high-resolution finite-element reference, residual near the corner, and loss versus wall-clock time. Success means lower error at a fixed training budget, particularly when boundary conditions or coefficients change so that the fixed power map is not optimal.

Verification

Beats tuned baseline

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Confirmed ✓

Stage 2 · Practical benchmark result: Significant win ✓

Methodology: 8 paired seeds, permutation test (p<0.05) against a learning-rate-tuned baseline. How verification works

Stage 1 — Mechanism check agent confidence 8/10

Built a readable MVP for positive-density monotone radial warps, with normalized cumulative integration, exponent and interpolation-convergence sweeps, and a small learned-warp neural regression comparison. The mechanism manifested: fitted r(s) exponents matched q, transformed singular exponents matched qλ within about 0.02%, and interpolation error slopes matched the predicted -qλ values within about 0.015. The learned warp slightly improved held-out RMSE over the fixed s² warp (0.00430 vs 0.00465), but required substantially more time and this is not evidence of a PDE-level win.

Agent confidence: 8/10

Baseline: Identity warp RMSE 0.02543 in 500 steps, 1.45 s; fixed r=s² warp RMSE 0.00465, 1.49 s.
Idea: Learned positive-density warp RMSE 0.00430 in 500 steps, 7.10 s; exponent law observed qλ: qλ=0.25, 0.50, 0.75 gave convergence slopes -0.255, -0.510, -0.765 versus predictions -0.25, -0.50, -0.75; exponent-fit errors were below roughly 0.02%.

Limitations: The mini-experiment is a 1D singular-function regression rather than the proposed 2D L-shaped Poisson Deep Ritz/PINN problem. It does not test PDE energy, polar Jacobian derivative transforms, boundary conditions, changing coefficients, GPU execution, or an outer unrolled optimization with a held-out residual; finite-difference derivative relative errors near vanishing densities are ill-conditioned.

How to run: python3 experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2607.23229) Sep 2, 2026

Beats the tuned baseline by 8.4% (p=0.0147, wins 7 / 8 paired seeds; metric: lower is better).

Benchmark: Poisson–Dirichlet process
Model: mlp_med
Paired seeds: 8
Baseline mean: 0.0002
Idea mean: 0.0002
Effect (Δ): -0 (−8.4%; negative = idea better)
Wins: 7 / 8 paired seeds
p-value: 0.0147 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±5.7%
Mechanism: Confirmed ✓
Practical effect: Helps
Baseline tuning: swept over 3 configs

Limitations:

The registered poisson_dirichlet track is a 1-D forcing-to-solution Dirichlet regression benchmark rather than the proposed full 2-D L-shaped Deep Ritz/PINN with explicit transformed derivatives, Jacobian-weighted energy, and unrolled outer optimization. The warp is applied to the predicted spatial output grid rather than implemented as a full polar PDE coordinate transform. Only three learning rates and 25 training epochs were tested; no larger models, longer budgets, fixed s^2 comparator, or coefficient/boundary-condition shifts were evaluated.

How to run: python3 bench_warp_registered.py

Verdict computed by deterministic test code from paired-seed statistics — not by the language model.

Stage 2 — Benchmark Aug 31, 2026 latest Worked ✓

Agent confidence: 8/10

On the registered poisson_dirichlet PDE track, the learned positive monotone radial warp improved test MSE over the identically sized raw-coordinate MLP. The best idea setting (lr=0.003, 25 epochs) achieved mean MSE 0.0002235653 versus the tuned baseline mean 0.0002440518 across 8 paired seeds, with delta_mean=-2.04866e-05 and permutation p=0.01475; this satisfies the harness criterion for a significant win. The trained-model signature also confirmed finite trained behaviour and lower spatial roughness for the warped system, although the signature is qualitative rather than a direct PDE exponent fit.

Baseline: poisson_dirichlet, mlp_med, tuned lr=0.003 and 25 epochs; 8-seed mean MSE 0.00024405181284237187, std 1.798621374535186e-05.
Idea: poisson_dirichlet, same field architecture plus learned positive normalized monotone warp with q=2; lr=0.003 and 25 epochs; 8-seed mean MSE 0.00022356524641509168, std 1.1563454829705651e-05.
Paired delta: -2.0486566427280195e-05; idea wins 7/8 paired seeds; permutation p=0.01475.

Benchmark result