# Quadratic-anisotropy electrostatic surrogate

- ID: 1316
- Canonical URL: https://synthcore.org/idea/1316/quadratic-anisotropy-electrostatic-surrogate
- API JSON: https://synthcore.org/api/idea/1316.json
- API Markdown: https://synthcore.org/api/idea/1316.md
- Verification status: unverified
- Source: [arXiv:2607.23315](https://arxiv.org/abs/2607.23315)
- Category: training
- Solves: speedup, scalability, stability
- ML areas: loss, training, inference-speedup
- Math tags: pde, convex-analysis, functional-analysis, optimization
- Ratings: usefulness 6/10; difficulty 4/10; novelty 7/10

## Idea description

Use the small dielectric-anisotropy estimate to replace expensive nonlocal electrostatic solves with a local field-energy surrogate during most neural-network updates. Periodically evaluate the exact field, estimate the approximation constant, and trigger correction solves only when the observed error exceeds the predicted quadratic scale.

## Mathematical statement

Let \(\delta\) be the dielectric anisotropy and write \(\varepsilon(\mathbf n)=\varepsilon_\perp I+\delta\,\mathbf n\otimes\mathbf n\), where \(\varepsilon_\perp>0\), \(\mathbf n\) is a unit director, and \(I\) is the identity. Let \(\mathcal E_{\rm exact}(\mathbf n)\) denote the minimized nonlocal electrostatic energy obtained by solving Gauss' law with the anisotropic tensor, and let \(\mathcal E_{\rm loc}(\mathbf n)\) denote the local approximation that neglects the induced depolarization or back-reaction field. The paper establishes a uniform quadratic estimate of the form \(|\mathcal E_{\rm exact}(\mathbf n)-\mathcal E_{\rm loc}(\mathbf n)|\le C\delta^2\), with \(C\) independent of the admissible director configuration under the paper's boundedness assumptions. The useful property is first-order cancellation: the local approximation has no \(O(\delta)\) energy error. A practical neural loss is the local energy plus an optional detached uncertainty margin proportional to \(C_{\rm est}\delta^2\), where \(C_{\rm est}\) is estimated from occasional exact solves.

## Key formulas

- $$\varepsilon(\mathbf n)=\varepsilon_\perp I+\delta\,\mathbf n\otimes\mathbf n,$$
- $$\left|\mathcal E_{\rm exact}(\mathbf n)-\mathcal E_{\rm loc}(\mathbf n)\right|\le C\delta^2,$$
- $$\mathcal L_{\rm cheap}(\theta)=\mathcal L_{\rm elastic}(\theta)+\mathcal E_{\rm loc}(\mathbf n_\theta)+\lambda C_{\rm est}\delta^2,$$
- $$C_{\rm est}\leftarrow\max\left(C_{\rm est},\frac{|\mathcal E_{\rm exact}(\mathbf n_\theta)-\mathcal E_{\rm loc}(\mathbf n_\theta)|}{\delta^2+10^{-12}}\right).$$

## Implementation notes

Use this in a PINN, neural operator, or differentiable simulator that repeatedly evaluates an electrostatic energy for a predicted material field. At every iteration, produce `n_theta`, compute the elastic loss, and evaluate the cheap electrostatic branch using a fixed applied field `E0`, for example `E_loc = 0.5*mean(E0.T @ eps(n) @ E0)`, or use the isotropic Poisson solution if the problem is charge-driven. Do not solve the anisotropic Poisson equation on every batch. Every `M=50–200` updates, freeze `n_theta`, solve the exact anisotropic field problem using either the primal potential or the Hodge-dual flux method, and calculate `gap = abs(E_exact-E_loc)`. Update `C_est` with the displayed ratio. Use the cheap loss alone, or add the detached margin `lambda*C_est*delta**2`; alternatively trigger exact correction steps when `gap > tau*C_est*delta**2`. The theorem supplies the quadratic dependence, while `C_est`, `M`, `tau`, and `lambda` are empirical. Start on a 2-D square with a small MLP and compare against exact-solve-every-step training. Sweep `delta` over `0.01`, `0.03`, and `0.1`; verify that the measured energy gap scales like `delta^2`. Then measure expensive-solve count, wall-clock training time, final exact energy, and field error. Success is 3–10x fewer PDE solves with no meaningful degradation at small anisotropy.

## Disclaimer

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