✓✓ Beats tuned baseline 2026

Hodge-dual electrostatic loss

Implementation & benchmark of arXiv:2607.23315 — Variational principles for the interaction of liquid crystals and electric fields in the Oseen--Frank model

Usefulness7/10

Difficulty6/10

Novelty8/10

Optimization Stability Speedup Accuracy Loss Training Architecture Convex analysis Geometry Pde Optimization

Source paper: Variational principles for the interaction of liquid crystals and electric fields in the Oseen--Frank model arXiv:2607.23315 ⓘ · analyzed Aug 30, 2026

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 Not peer-reviewed.

Idea description

Replace a jointly optimized scalar electrostatic potential in a neural PDE solver with a dual flux represented by a Hodge curl correction. The resulting inner problem is a positive quadratic minimization with the divergence constraint satisfied exactly, avoiding unstable primal-dual training dynamics.

Formulas

$$2W_e(\widetilde{\mathbf n})=K_1(\operatorname{div}\widetilde{\mathbf n})^2+K_2(\widetilde{\mathbf n}\cdot\operatorname{curl}\widetilde{\mathbf n})^2+K_3|\widetilde{\mathbf n}\times\operatorname{curl}\widetilde{\mathbf n}|^2,$$

$$W_e(\widetilde{\mathbf n})=\frac K2|\nabla\widetilde{\mathbf n}|^2\qquad(K_1=K_2=K_3=K),$$

$$\min_{\mathbf p:\,\operatorname{div}\mathbf p=\rho}\frac12\int_\Omega\mathbf p^T\varepsilon(\mathbf n)^{-1}\mathbf p\,dx,$$

$$\mathbf p=\mathbf p_0+\operatorname{curl}\mathbf A,\qquad \mathcal L(\theta,\psi)=\frac K2\int_\Omega|\nabla\mathbf n_\theta|^2dx+\frac12\int_\Omega(\mathbf p_0+\operatorname{curl}\mathbf A_\psi)^T\varepsilon(\mathbf n_\theta)^{-1}(\mathbf p_0+\operatorname{curl}\mathbf A_\psi)dx.$$

Mathematical statement

The paper gives the one-constant elastic density $W_e(\widetilde{\mathbf n})=\frac K2|\nabla\widetilde{\mathbf n}|^2$, obtained from $2W_e=K_1(\operatorname{div}\widetilde{\mathbf n})^2+K_2(\widetilde{\mathbf n}\cdot\operatorname{curl}\widetilde{\mathbf n})^2+K_3|\widetilde{\mathbf n}\times\operatorname{curl}\widetilde{\mathbf n}|^2$ when $K_1=K_2=K_3=K$. Let $\mathbf n:\Omega\to\mathbb R^3$ be the predicted unit director, $\varepsilon(\mathbf n)\succ0$ the dielectric tensor, $\rho$ the charge density, $\phi$ the scalar potential, and $\mathbf p=\varepsilon(\mathbf n)\nabla\phi$ the electric displacement. Convex duality transforms the electrostatic solve into $\min_{\mathbf p:\operatorname{div}\mathbf p=\rho}\frac12\int_\Omega\mathbf p^T\varepsilon(\mathbf n)^{-1}\mathbf p\,dx$. Choose a reference flux $\mathbf p_0$ with $\operatorname{div}\mathbf p_0=\rho$, and parameterize $\mathbf p=\mathbf p_0+\operatorname{curl}\mathbf A$; since $\operatorname{div}(\operatorname{curl}\mathbf A)=0$, Gauss' law is automatic. Positive definiteness of $\varepsilon$ makes the flux minimizer unique, apart from vector-potential gauge freedom.

Implementation notes

Integrate this into a neural PDE solver or differentiable field simulator whose network currently predicts both a material field and a scalar electrostatic potential. Let n_theta(x) output three channels and normalize them pointwise as n = raw_n / (norm(raw_n) + 1e-8). Let A_psi(x) output three vector-potential channels. On a regular 2-D or 3-D grid, calculate spatial derivatives with finite differences or automatic differentiation, compute curl(A_psi), and form p = p0 + curl(A_psi). Construct p0 once using a constant-coefficient Poisson solve or an analytic flux satisfying div(p0)=rho. Use eps(n) = eps_perp*I + eps_a*n*n^T, requiring eps_perp > abs(eps_a), and evaluate 0.5 * p.T @ inverse(eps(n)) @ p at quadrature points. Add the elastic loss 0.5*K*|grad(n)|^2 and any boundary penalties. Alternate 1–5 Adam updates of psi with one update of theta, or differentiate through 10–30 conjugate-gradient iterations for the inner quadratic solve. The curl parameterization is computed exactly; only p0, derivatives, and quadrature are numerical approximations. First test on a unit square or cube with manufactured smooth rho and known dielectric tensor. Compare against a primal PINN that alternates optimization of n and phi. Record Gauss-law residual, total energy, gradient variance, failed runs, and wall-clock time to a fixed energy tolerance. Success is lower residual and fewer unstable runs at equal network evaluations.

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 periodic-grid Hodge-dual electrostatic loss with p=p0+curl(A), anisotropic positive dielectric energy, and exact discrete divergence preservation. The mechanism checks passed: div(curl A) was <=7.45e-9 even at amplitude 1e4, lambda_max*epsilon had 0% spread across epsilon={0.5,1,2,4}, and the gradient-descent transition occurred near the predicted eta*lambda_max=2 boundary. The toy primal and dual quadratic problems had identical nonzero condition numbers, so no universal speedup was observed.

Agent confidence: 8/10

Baseline: Primal/dual constant-coefficient condition number: 29.8564; no measured conditioning or speed advantage.
Idea: Maximum discrete Gauss residual 7.45e-9; epsilon*lambda_max=[2,2,2,2], relative spread 0%; GD final/initial energy ratios at eta/critical=[0.9,0.99,1.01,1.1] were [2.39e-6,0.310,3.154,39130.2].

Limitations: Only a small 2-D periodic finite-difference toy was tested, with constant dielectric scaling and a hand-constructed flux/potential rather than trained neural networks. Boundary conditions, nonconstant anisotropic dielectric fields, manufactured nonzero charge solves, wall-clock comparisons, alternating theta/psi optimization, and CUDA behavior were not tested.

How to run: python3 run_experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2607.23315) Sep 2, 2026

Beats the tuned baseline by 62.9% (p=0.0081, wins 8 / 8 paired seeds; metric: lower is better).

Benchmark: Poisson_hodge_periodic
Model: mlp_tiny
Paired seeds: 8
Baseline mean: 1.0801
Idea mean: 0.4011
Effect (Δ): -0.6789 (−62.9%; negative = idea better)
Wins: 8 / 8 paired seeds
p-value: 0.0081 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±6.8%
Mechanism: Confirmed ✓
Practical effect: Helps
Baseline tuning: swept over 3 configs

Limitations:

The benchmark uses a small periodic 2-D manufactured Poisson task with constant diagonal dielectric and a basis-based evaluation reconstruction; it does not test physical boundary conditions, nonconstant anisotropic dielectric tensors, learned director fields, alternating theta/psi updates, wall-clock speed at scale, or the built-in tabular/vision/sequence/dynamics tracks.

How to run: python3 bench_stage2.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

Implemented an end-to-end periodic Poisson benchmark with identical MLP systems: the baseline trains a scalar potential using the primal variational objective, while the idea trains a vector potential and forms p=p0+curl(A) with the positive dual energy. Across 8 paired seeds, the idea achieved MSE 0.4011 versus baseline 1.0801, delta=-0.6789, with all 8 wins and permutation p=0.0081; the trained-model mechanism signature confirmed max |div(curl A)|=2.38e-7. This is a real benchmark win on the matched PDE task, though it does not establish a general speedup or superiority under boundary/nonconstant-dielectric settings.

Baseline: map[best_cfg:map[lr:0.003] full_mean:1.0800504088401794 full_std:0.10522868591444227 per_seed:[1.1444686651229858 0.8914604783058167 1.1053142547607422 1.2771755456924438 1.091360092163086 1.0986981391906738 1.034430980682373 0.9974951148033142] sweep:[map[cfg:map[lr:0.001] mean:2.206587553024292] map[cfg:map[lr:0.003] mean:1.1046047359704971] map[cfg:map[lr:0.01] mean:1.2606632113456726]]]
Comparison: map[delta_mean:-0.6789125688374043 idea_wins:8 mde:0.07381216084595137 mde_rel_pct:6.834140355098345 n_pairs:8 p_value:0.0081 verdict:idea better (significant)]
Idea: map[full_mean:0.4011378400027752 full_std:0.059436192075174005 per_seed:[0.33402565121650696 0.3543514013290405 0.42588725686073303 0.5025325417518616 0.3978813886642456 0.43913376331329346 0.4413507878780365 0.3139399290084839] selected_cfg:map[lr:0.01]]
Mechanism signature: map[baseline_observed_constraint_residual:106.66156768798828 confirmed:true idea_observed_div_curl_max:2.384185791015625e-07 prediction:div(curl A)=0 for the trained dual model]

Benchmark result