# Derivative-conditioned fine-scale corrector

- ID: 1850
- Canonical URL: https://synthcore.org/idea/1850/derivative-conditioned-fine-scale-corrector
- API JSON: https://synthcore.org/api/idea/1850.json
- API Markdown: https://synthcore.org/api/idea/1850.md
- Verification status: beats_baseline
- Source: [arXiv:2608.05580](https://arxiv.org/abs/2608.05580)
- Category: architecture
- Solves: speedup, accuracy, memory
- ML areas: cnn, memory, inference-speedup
- Math tags: approximation-theory, multiscale-analysis, pde, tensor-decomposition
- Ratings: usefulness 8/10; difficulty 5/10; novelty 7/10

## Idea description

Replace a full-resolution neural operator with an explicit multiscale reconstruction: a coarse predictor produces a low-resolution field, and a small corrector module combines its gradients and Hessians with learned rapidly varying basis functions. The model should recover fine detail without running the expensive backbone at fine resolution.

## Mathematical statement

The paper's second-order asymptotic expansion represents a fine-scale field using macroscopic derivatives and local correctors. For moisture, the extracted formula is \(\omega^{(2)}(\mathbf{x},\mathbf{y},t)=\mathcal{J}_{\alpha_{1}\alpha_{2}}(\mathbf{y},\omega^{(0)})\frac{\partial^{2}\omega^{(0)}(\mathbf{x},t)}{\partial x_{\alpha_{1}}\partial x_{\alpha_{2}}}+\mathcal{I}_{\alpha_{1}}(\mathbf{y},\omega^{(0)})\frac{\partial\omega^{(0)}(\mathbf{x},t)}{\partial x_{\alpha_{1}}}-\mathcal{F}_{\alpha_{1}\alpha_{2}}(\mathbf{y},\omega^{(0)})\frac{\partial\omega^{(0)}}{\partial x_{\alpha_{1}}}\frac{\partial\omega^{(0)}}{\partial x_{\alpha_{2}}}-\mathbb{S}(\mathbf{y},T^{(0)})\). Here \(\mathbf{x}\) is the coarse coordinate, \(\mathbf{y}=\mathbf{x}/\epsilon\) is the fast within-cell coordinate, \(\epsilon\) is the fine-to-coarse scale ratio, \(\omega^{(0)}\) is the homogenized field, \(\mathcal{J}\) and \(\mathcal{I}\) are second- and first-derivative correctors, \(\mathcal{F}\) multiplies quadratic gradient interactions, and \(\mathbb{S}\) is a source corrector. The same structure appears for displacement through tensors \(\mathcal{X}_{ih}^{\alpha_1\alpha_2}\) and \(\mathcal{Q}_{ih}^{\alpha_1}\). The reconstruction uses explicit \(\epsilon\) and \(\epsilon^2\) scaling, so the model is not merely adding an unconstrained residual branch.

## Key formulas

- $$\omega^{(2)}(\mathbf{x},\mathbf{y},t)=\mathcal{J}_{\alpha_{1}\alpha_{2}}(\mathbf{y},\omega^{(0)})\frac{\partial^{2}\omega^{(0)}(\mathbf{x},t)}{\partial x_{\alpha_{1}}\partial x_{\alpha_{2}}}+\mathcal{I}_{\alpha_{1}}(\mathbf{y},\omega^{(0)})\frac{\partial\omega^{(0)}(\mathbf{x},t)}{\partial x_{\alpha_{1}}}-\mathcal{F}_{\alpha_{1}\alpha_{2}}(\mathbf{y},\omega^{(0)})\frac{\partial\omega^{(0)}}{\partial x_{\alpha_{1}}}\frac{\partial\omega^{(0)}}{\partial x_{\alpha_{2}}}-\mathbb{S}(\mathbf{y},T^{(0)})$$
- $$\omega_{\Delta}^{(2,\epsilon)}(\mathbf{x},t)=-\epsilon\mathcal{J}_{\alpha_{1}}\frac{\partial\omega^{(0)}}{\partial x_{\alpha_{1}}}-\epsilon^{2}\left(\mathcal{J}_{\alpha_{1}\alpha_{2}}\frac{\partial^{2}\omega^{(0)}}{\partial x_{\alpha_{1}}\partial x_{\alpha_{2}}}+\mathcal{I}_{\alpha_{1}}\frac{\partial\omega^{(0)}}{\partial x_{\alpha_{1}}}-\mathcal{F}_{\alpha_{1}\alpha_{2}}\frac{\partial\omega^{(0)}}{\partial x_{\alpha_{1}}}\frac{\partial\omega^{(0)}}{\partial x_{\alpha_{2}}}-\mathbb{S}\right)$$
- $$\widehat{u}_{\mathrm{fine}}(\mathbf{x},\mathbf{y})=u^{(0)}(\mathbf{x})+\epsilon\,\mathcal{X}^{\alpha}_{h}(\mathbf{y},z)\,\partial_{\alpha}u_h^{(0)}(\mathbf{x})+\epsilon^{2}\left[\mathcal{X}^{\alpha\beta}_{h}(\mathbf{y},z)\,\partial_{\alpha\beta}u_h^{(0)}(\mathbf{x})+\mathcal{Q}^{\alpha}_{h}(\mathbf{y},z)\,\partial_{\alpha}u_h^{(0)}(\mathbf{x})-\mathcal{F}^{\alpha\beta}(\mathbf{y},z)\,\partial_{\alpha}u^{(0)}(\mathbf{x})\partial_{\beta}u^{(0)}(\mathbf{x})\right]$$

## Implementation notes

(1) Integration point: use a 2D CNN or neural operator whose backbone predicts a coarse field \(u^{(0)}\in\mathbb{R}^{H/r\times W/r\times C}\), followed by factor-\(r\) interpolation and a fine-scale corrector head. The head receives the within-cell coordinate \(\mathbf{y}\), the coarse field, finite-difference gradients, and finite-difference Hessians. It outputs corrector tensors or their coefficients at each fine pixel. (2) Pseudocode: `u0 = coarse_net(x_coarse); g = grad(u0); H = hessian(u0); z = state_encoder(u0, material_context); basis = corrector_net(y_grid, z); residual = eps*contract(basis.J1,g) + eps**2*(contract(basis.J2,H) + contract(basis.I1,g) - contract(basis.F, outer(g,g)) - basis.S); uhat = interpolate(u0) + residual`. Symmetrize the Hessian and use reflective or periodic finite differences. (3) Compute the derivative structure and \(\epsilon,\epsilon^2\) scaling exactly from the asymptotic formula; learn the basis tensors by gradient descent. If the fine material pattern is known, provide a local patch or Fourier encoding of \(\mathbf{y}\). Train with field, gradient, and optional PDE-residual losses. (4) First experiment: use 2D heterogeneous diffusion or elasticity data, comparing against a full-resolution U-Net or Fourier Neural Operator at equal parameter count and FLOPs. Success is at least 2x lower inference FLOPs or 2x smaller activation memory at matched relative \(L^2\) error, especially on high-frequency Fourier modes. Ablate Hessian and quadratic-gradient terms to verify that the asymptotic structure, rather than extra parameters alone, produces the improvement.

## Verification

- Status: beats_baseline
- Mechanism evidence: yes
- Mechanism confirmed: no
- Practical verdict: helps
- Verdict: Built a compact derivative-conditioned multiscale corrector with periodic Fourier upsampling, exact gradient/Hessian features, local fast-coordinate bases, and explicit epsilon/epsilon^2 scaling. The mechanism manifested: measured first- and second-order correction slopes were 1.00 and 2.00 versus predictions of 1 and 2; setting epsilon to zero removed the correction exactly, and removing Hessian features increased relative error from 0.0178 to 0.4023. On this synthetic, coefficient-matched task the structured corrector reduced relative L2 error from 0.7360 for interpolation to 0.0183, but this is not evidence of a general neural-operator win because the data-generating correctors were deliberately within the fitted basis.

### Mechanism check

- Verdict: Built a compact derivative-conditioned multiscale corrector with periodic Fourier upsampling, exact gradient/Hessian features, local fast-coordinate bases, and explicit epsilon/epsilon^2 scaling. The mechanism manifested: measured first- and second-order correction slopes were 1.00 and 2.00 versus predictions of 1 and 2; setting epsilon to zero removed the correction exactly, and removing Hessian features increased relative error from 0.0178 to 0.4023. On this synthetic, coefficient-matched task the structured corrector reduced relative L2 error from 0.7360 for interpolation to 0.0183, but this is not evidence of a general neural-operator win because the data-generating correctors were deliberately within the fitted basis.
- Confidence: 8/10
- Limitations: No CNN or neural-operator training, wall-clock/FLOP/activation-memory measurement, heterogeneous PDE solver data, material-conditioned basis learning, noisy or out-of-distribution tests, or equal-parameter baseline was evaluated. Sample-count scaling was flat because the synthetic corrector coefficients were exactly identifiable from one noiseless sample.

### 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/657/bench_report.json)
- [bench_stage2.py](https://synthcore.org/code/657/bench_stage2.py)
- [custom_pde_track.py](https://synthcore.org/code/657/custom_pde_track.py)
- [experiment.py](https://synthcore.org/code/657/experiment.py)
- [report.md](https://synthcore.org/code/657/report.md)
- [report_bench_2026-08-31T212025.md](https://synthcore.org/code/657/report_bench_2026-08-31T212025.md)
- [results.json](https://synthcore.org/code/657/results.json)
- [Download all files as ZIP](https://synthcore.org/download/657)

## Disclaimer

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