# Corrector-Enriched Two-Scale Network

- ID: 976
- Canonical URL: https://synthcore.org/idea/976/corrector-enriched-two-scale-network
- API JSON: https://synthcore.org/api/idea/976.json
- API Markdown: https://synthcore.org/api/idea/976.md
- Verification status: beats_baseline
- Source: [arXiv:2607.15702](https://arxiv.org/abs/2607.15702)
- Category: architecture
- Solves: accuracy, sample-efficiency, stability
- ML areas: mlp, pde, initialization, sample-efficiency
- Math tags: pde, approximation-theory, numerical-analysis, homogenization
- Ratings: usefulness 8/10; difficulty 5/10; novelty 6/10

## Idea description

Replace a single neural representation of a rapidly oscillatory PDE solution by a macroscopic network plus an explicitly oscillatory corrector network. Feed the network both the slow coordinate $x$ and fast coordinate $y=x/\varepsilon$, and train the resulting composite field in a variational energy objective. This targets the paper's scale-robust approximation bound rather than forcing the optimizer and finite sample set to discover oscillations of wavelength $\varepsilon$.

## Mathematical statement

The paper states the two-scale approximation estimate $\mathcal{A}_{m}^{\varepsilon}\leq C\bigl(\varepsilon+\Phi_{0,m_{0}}^{2}+\Phi_{1,m_{1}}^{2}\bigr)$. Here $\mathcal{A}_{m}^{\varepsilon}$ is the population approximation error at microscopic scale $\varepsilon$, $C$ is independent of $\varepsilon$, $m=m_0+m_1$ is the total feature budget, and $\Phi_{0,m_0}$ and $\Phi_{1,m_1}$ are approximation errors of the coarse and first-order corrector feature classes with budgets $m_0$ and $m_1$. The transferable property is that the irreducible homogenization error is $O(\varepsilon)$ while the learned coarse and oscillatory components are approximated separately. Use the adapted ansatz $u_\theta^\varepsilon(x)=u_{0,\theta_0}(x)+\varepsilon\sum_{j=1}^{d}\chi_{j,\theta_1}(x,x/\varepsilon)\,\partial_j u_{0,\theta_0}(x)$, where $u_0$ is the macroscopic network, $\chi_j$ are learned or precomputed periodic correctors, $d$ is spatial dimension, and $\partial_j u_0$ is the coarse derivative.

## Key formulas

- $$\mathcal{A}_{m}^{\varepsilon}\leq C\bigl(\varepsilon+\Phi_{0,m_{0}}^{2}+\Phi_{1,m_{1}}^{2}\bigr)$$
- $$u_\theta^\varepsilon(x)=u_{0,\theta_0}(x)+\varepsilon\sum_{j=1}^{d}\chi_{j,\theta_1}\!\left(x,\frac{x}{\varepsilon}\right)\partial_j u_{0,\theta_0}(x)$$
- $$\mathcal{L}_{\mathrm{var}}(\theta)=\frac{1}{N}\sum_{n=1}^{N}\left[W\!\left(x_n,\nabla u_\theta^\varepsilon(x_n)\right)-f(x_n)u_\theta^\varepsilon(x_n)\right]$$

## Implementation notes

1. Integration point: modify the solution MLP used by a variational PINN or Deep Ritz solver. Keep a coarse MLP $u_{0,\theta_0}(x)$, add a corrector MLP $\chi_{\theta_1}(x,y)$, and evaluate it at $y=x/\varepsilon$; enforce homogeneous boundary conditions by multiplying the final output by a known boundary mask $b(x)$ or by adding a boundary extension $g(x)$. Use automatic differentiation to obtain $\nabla_x u_\theta^\varepsilon(x)$, including the chain-rule derivative through $y=x/\varepsilon$. 2. Pseudocode: sample collocation points $x_n$; set $y_n=x_n/\varepsilon$; compute $u_0=u_0(x_n)$ and $J_0=\nabla_xu_0(x_n)$; compute $\chi=\chi(x_n,y_n)$; form $u=u_0+\varepsilon\sum_j\chi_jJ_{0,j}$; compute $L=N^{-1}\sum_n[W(x_n,\nabla u_n)-f_nu_n]$; backpropagate and update both parameter sets. 3. The theoretical objects are the coarse and first-order feature errors $\Phi_{0,m_0}$ and $\Phi_{1,m_1}$; estimate them empirically using held-out energy error and gradient error. If coefficient correctors are known, initialize or freeze $\chi_j$ from cell-problem data; otherwise pretrain them on periodic cell residuals and enforce periodicity with Fourier features or modulo-one coordinates. 4. First experiment: solve a 2D nonlinear periodic elliptic equation at $\varepsilon\in\{1/8,1/16,1/32\}$ using the same MLP width and the same number of quadrature points. Compare a plain coordinate MLP against the two-scale ansatz at equal parameter count and equal FLOPs. Measure relative $H^1$ error, energy error, gradient variance, and updates to a fixed error. Success means error scaling remains roughly proportional to $\varepsilon$ plus feature error, while the plain MLP loses accuracy or requires substantially more samples as $\varepsilon$ shrinks.

## Verification

- Status: beats_baseline
- Mechanism evidence: yes
- Mechanism confirmed: yes
- Practical verdict: helps
- Verdict: Built a variational 1D two-scale network benchmark with a coarse MLP and explicitly supplied periodic corrector, plus a plain coordinate MLP baseline. The mechanism checks passed: truncation H1 error scaled as epsilon^1.001, consistent with the predicted O(epsilon) behavior, and squared H1 error scaled as (corrector-amplitude mismatch)^1.000, equivalent to linear H1 error versus amplitude mismatch. At equal parameter count, the two-scale model substantially beat the plain MLP across all tested epsilon values, although the corrector was precomputed rather than learned.

### Mechanism check

- Verdict: Built a variational 1D two-scale network benchmark with a coarse MLP and explicitly supplied periodic corrector, plus a plain coordinate MLP baseline. The mechanism checks passed: truncation H1 error scaled as epsilon^1.001, consistent with the predicted O(epsilon) behavior, and squared H1 error scaled as (corrector-amplitude mismatch)^1.000, equivalent to linear H1 error versus amplitude mismatch. At equal parameter count, the two-scale model substantially beat the plain MLP across all tested epsilon values, although the corrector was precomputed rather than learned.
- Confidence: 8/10
- Limitations: Only a 1D manufactured elliptic problem was tested, not the requested 2D nonlinear PDE. The corrector was known and frozen, so this validates the precomputed-corrector variant rather than learned-corrector optimization. Derivative/FLOP costs were not matched, only parameter count and training sample/step budget; the epsilon law is for weighted H1 truncation error rather than a full population approximation-theorem bound.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 0.0081
- Paired wins: 8/8
- Benchmark verdict: idea better (significant)

## Artifacts

- [bench_report.json](https://synthcore.org/code/174/bench_report.json)
- [bench_two_scale.py](https://synthcore.org/code/174/bench_two_scale.py)
- [experiment.py](https://synthcore.org/code/174/experiment.py)
- [report.md](https://synthcore.org/code/174/report.md)
- [report_bench_2026-08-30T193324.md](https://synthcore.org/code/174/report_bench_2026-08-30T193324.md)
- [results.json](https://synthcore.org/code/174/results.json)
- [Download all files as ZIP](https://synthcore.org/download/174)

## Disclaimer

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