# Offline-online conditional corrector bank

- ID: 1851
- Canonical URL: https://synthcore.org/idea/1851/offline-online-conditional-corrector-bank
- API JSON: https://synthcore.org/api/idea/1851.json
- API Markdown: https://synthcore.org/api/idea/1851.md
- Verification status: mech_ok_no_baseline
- Source: [arXiv:2608.05580](https://arxiv.org/abs/2608.05580)
- Category: memory
- Solves: speedup, scalability, accuracy
- ML areas: memory, inference-speedup, world-model
- Math tags: multiscale-analysis, tensor-decomposition, approximation-theory, numerical-analysis
- Ratings: usefulness 7/10; difficulty 6/10; novelty 8/10

## Idea description

Build a reusable bank of local fine-scale correctors offline, then let a lightweight online network assemble them using the current coarse state and material context. This replaces repeatedly applying a globally expensive fine-scale model with low-dimensional coefficient modulation for parameter sweeps and autoregressive rollout.

## Mathematical statement

The paper replaces rapidly oscillating coefficients such as \(k_{ij}^{\epsilon}(\mathbf{x},T^{\epsilon})\) and \(C_{ijkl}^{\epsilon}(\mathbf{x},T^{\epsilon})\) by effective coefficients \(\hat{k}_{ij}(T^{(0)})\) and \(\hat{C}_{ijkl}(T^{(0)})\), while higher-order fields use local tensors \(\mathcal{J},\mathcal{I},\mathcal{F},\mathcal{X},\mathcal{Q}\) evaluated at the fast coordinate and coarse state. This creates an offline-online separation: expensive local response objects are computed once, and each online query solves or predicts only the coarse state and a small set of coefficients. In the neural adaptation, \(b_m(\mathbf{y})\) are fixed local basis functions, \(a_m(z)\) are online coefficients, \(z\) is a low-dimensional embedding of coarse fields and parameters, and \(\phi_m\) are derivative features. The reconstruction is low-rank in the fast coordinate, while state dependence is represented through the coefficients.

## Key formulas

- $$k_{ij}^{\epsilon}(\mathbf{x},T^{\epsilon})\ \longrightarrow\ \hat{k}_{ij}(T^{(0)}),\qquad C_{ijkl}^{\epsilon}(\mathbf{x},T^{\epsilon})\ \longrightarrow\ \hat{C}_{ijkl}(T^{(0)})$$
- $$\omega^{(2)}(\mathbf{x},\mathbf{y},t)=\mathcal{J}_{\alpha_{1}\alpha_{2}}(\mathbf{y},\omega^{(0)})\partial_{\alpha_{1}\alpha_{2}}\omega^{(0)}+\mathcal{I}_{\alpha_{1}}(\mathbf{y},\omega^{(0)})\partial_{\alpha_{1}}\omega^{(0)}-\mathcal{F}_{\alpha_{1}\alpha_{2}}(\mathbf{y},\omega^{(0)})\partial_{\alpha_{1}}\omega^{(0)}\partial_{\alpha_{2}}\omega^{(0)}-\mathbb{S}(\mathbf{y},T^{(0)})$$
- $$u_{\mathrm{fine}}(\mathbf{x},\mathbf{y})=\operatorname{interp}(u^{(0)})(\mathbf{x},\mathbf{y})+\sum_{m=1}^{M}a_m(z(\mathbf{x}))\,b_m(\mathbf{y})\,\phi_m\bigl(u^{(0)}(\mathbf{x}),\nabla u^{(0)}(\mathbf{x}),\nabla^2u^{(0)}(\mathbf{x})\bigr)$$
- $$\mathcal{L}_{\mathrm{total}}=\|u_{\mathrm{fine}}-u_{\mathrm{target}}\|_2^2+\lambda_g\|\nabla u_{\mathrm{fine}}-\nabla u_{\mathrm{target}}\|_2^2+\lambda_r\|\mathcal{R}(u_{\mathrm{fine}})\|_2^2+\lambda_o\sum_{m=1}^{M}\|b_m\|_2^2$$

## Implementation notes

(1) Integration point: insert the module after the coarse representation in a neural operator, diffusion denoiser, or recurrent world-model transition. The offline component is a compact basis bank over local coordinates and material/context bins; the online component is an MLP mapping pooled coarse features and conditioning variables to coefficients \(a_m(z)\). (2) Pseudocode: `# offline: r = target - interpolate(coarse); compute randomized SVD r ~= sum_m b[m](y)*c[m]; normalize b; # online: u0 = coarse_net(input); z = MLP(pool(u0), parameters); g,H = finite_difference(u0); phi = concat(u0,g,H,outer(g,g)); a = coeff_net(z); rhat = sum_m a[:,m,None,None] * b[m] * phi[m]; uhat = interpolate(u0)+rhat`. Retain the smallest \(M\) singular vectors explaining 99% of residual energy, and impose zero correction for zero derivatives and zero source. (3) The SVD, basis normalization, and effective coarse coefficients are computed offline; the online network estimates only \(z\) and \(a_m\). Monitor the singular-value tail and coefficient smoothness across neighboring conditioning states. (4) First experiment: generate 2D diffusion or elasticity fields with random checkerboard or fiber microstructures and 100 parameter-query trajectories per structure. Compare a full-resolution U-Net/FNO, a coarse-only model, and the banked-corrector model. Measure amortized training-plus-inference cost, fine-grid relative error, and long-rollout error. The target is at least 3x lower online latency at matched error, with a small basis size and graceful interpolation when the conditioning parameter lies between training values.

## Verification

- Status: mech_ok_no_baseline
- Mechanism evidence: yes
- Mechanism confirmed: yes
- Practical verdict: no_effect
- Verdict: Built an offline SVD corrector bank with online context interpolation and derivative-feature gating in corrector_bank_experiment.py. The mechanism manifested: SVD projection error matched predicted omitted tail energy within 0.2 relative error, zero derivative features produced exactly zero correction, and halving context spacing reduced MSE by 15.2–17.4x, consistent with the predicted squared O(h^2) interpolation error. The bank reduced toy relative error from 1.0 for the coarse-only baseline to 3.13e-6 at M=4, but was slightly slower in this unoptimized Python toy, so no online speedup claim is supported.

### Mechanism check

- Verdict: Built an offline SVD corrector bank with online context interpolation and derivative-feature gating in corrector_bank_experiment.py. The mechanism manifested: SVD projection error matched predicted omitted tail energy within 0.2 relative error, zero derivative features produced exactly zero correction, and halving context spacing reduced MSE by 15.2–17.4x, consistent with the predicted squared O(h^2) interpolation error. The bank reduced toy relative error from 1.0 for the coarse-only baseline to 3.13e-6 at M=4, but was slightly slower in this unoptimized Python toy, so no online speedup claim is supported.
- Confidence: 8/10
- Limitations: Only a synthetic NumPy toy was tested: no 2D PDE solver, neural online MLP, full-resolution U-Net/FNO, autoregressive rollout, FLOP-normalized training, material structures, or realistic latency benchmark. The final interpolation sweep uses a controlled uniform offline context design and does not test sparse or noisy offline sampling.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 0.11105
- Paired wins: 3/8
- Benchmark verdict: no significant win

## Artifacts

- [bench_report.json](https://synthcore.org/code/687/bench_report.json)
- [corrector_bank_experiment.py](https://synthcore.org/code/687/corrector_bank_experiment.py)
- [pde_corrector_track.py](https://synthcore.org/code/687/pde_corrector_track.py)
- [report.md](https://synthcore.org/code/687/report.md)
- [report_bench_2026-08-31T215217.md](https://synthcore.org/code/687/report_bench_2026-08-31T215217.md)
- [results.json](https://synthcore.org/code/687/results.json)
- [run_bench.py](https://synthcore.org/code/687/run_bench.py)
- [Download all files as ZIP](https://synthcore.org/download/687)

## Disclaimer

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