Physics-informed neural decoder
Implementation & benchmark of arXiv:2607.06287 — Kernel-based Operator Learning: Error Analysis, Budget Allocation, and a Physics-Informed Extension
Source paper: Kernel-based Operator Learning: Error Analysis, Budget Allocation, and a Physics-Informed Extension arXiv:2607.06287 ⓘ · analyzed Aug 30, 2026
AI-generated research hypothesis, automatically tested. Not peer-reviewed.
Idea description
Attach an online PDE-constrained reconstruction layer to a neural operator instead of accepting its raw output field. The layer stays close to the network prediction at sampled output locations while minimizing a differentiable PDE residual and boundary-condition violation, allowing physics correction for each new input without retraining the neural operator.
Formulas
Mathematical statement
The paper factors the surrogate as \(\mathcal{A}=A_{\mathrm{on}}\circ A_{\mathrm{off}}\circ S_X\), where \(S_X\) samples the input, \(A_{\mathrm{off}}\) predicts a finite output-observation vector, and \(A_{\mathrm{on}}\) reconstructs a function. For an input \(u\), the exact discretized output is \(\boldsymbol{v}^{\dagger}=S_Y(\mathcal{G}(u))\), while the learned prediction is \(\widehat{\boldsymbol v}=A_{\mathrm{off}}(S_X(u))\); the decoder therefore receives perturbed data. The paper's physics-informed example uses \(-\Delta v(x)=u(x)\) on \(\Omega=(0,1)^2\), with \(v=0\) on \(\partial\Omega\). In the neural adaptation, \(f_\theta\) is a differentiable output-field parameterization, \(Y=\{y_j\}_{j=1}^m\) are data points, \(\widehat v_j\) are network predictions, \(Z=\{z_k\}_{k=1}^{m_{\mathcal L}}\) are PDE collocation points, \(Q=\{q_b\}_{b=1}^{m_\partial}\) are boundary points, \(\mathcal L\) is the PDE operator, and \(\mathcal B\) is the boundary operator. The reconstruction minimizes data mismatch plus PDE and boundary penalties.
Implementation notes
Integrate this as an optional inference-time decoder after an FNO, DeepONet, or operator transformer that outputs field values. For each new sample, take the network prediction \(\widehat v_j\) at output points \(Y\), the input function needed to evaluate the PDE right-hand side, and fixed collocation sets \(Z\subset\Omega\) and \(Q\subset\partial\Omega\). Represent the corrected field with a small coordinate MLP \(f_\theta(x)\), initialized by fitting \(f_\theta(y_j)\) to \(\widehat v_j\), or use \(f_\theta=\widehat v+g_\theta\). Run 10--100 Adam or L-BFGS steps on the displayed objective. Compute \(\mathcal Lf_\theta\) using automatic differentiation; for Poisson, evaluate the Laplacian with second derivatives. Enforce boundary conditions through the penalty or use \(f_\theta(x)=d(x)g_\theta(x)\), where \(d\) vanishes on the boundary. The mathematical terms are data fidelity, PDE residual, and boundary residual; \(\mu\), \(\beta\), and collocation locations are tuned on validation data. First test 2D Poisson with a small FNO baseline, comparing raw output, data-only reconstruction, and physics-informed reconstruction. Record relative \(L_2\) error, PDE residual, boundary error, and inference time. Success means lower error and residual at fixed training cost; saturation at the raw predictor error indicates the decoder has exposed the offline learning bottleneck.
Verification
Stage 1 · Toy mechanism gate: Passed ✓
Stage 2 · Mechanism transferred to benchmark: Confirmed ✓
Stage 2 · Practical benchmark result: Significantly worse ✗
Methodology: 8 paired seeds, permutation test (p<0.05) against a learning-rate-tuned baseline. How verification works
Stage 1 — Mechanism check agent confidence 9/10
Built an inference-time Poisson decoder using a Dirichlet sine basis, with an exact closed-form minimizer for the data-plus-PDE objective. The predicted contraction law 1/(1+mu*lambda^2), the penalty transition at mu*lambda^2=1, and stronger correction of higher-frequency modes were all confirmed to below 5e-16 absolute error; the large-penalty PDE limit reached 1.84e-7 relative error. In a synthetic imperfect-operator comparison, the decoder reduced relative L2 error and PDE residual substantially, but this is a mechanism-level MVP rather than evidence from a trained neural operator.
- Agent confidence
- 9/10
- Baseline
- Raw/data-only modal prediction: relative L2 0.2554, RMS PDE residual 9.299, boundary error 0.
- Idea
- At mu=0.01: relative L2 0.03084, RMS PDE residual 0.3328, boundary error 0; mechanism sweep max absolute prediction error 4.44e-16 and large-mu solution error 1.84e-7.
Limitations: The predictor was synthetic rather than a trained FNO/DeepONet, and the decoder used a spectral sine basis with a closed-form solve rather than a coordinate MLP with automatic second derivatives and 10-100 inference-time optimization steps. No wall-clock comparison against iterative autodiff decoding, inconsistent RHS robustness test, or full neural-operator training experiment was performed.
How to run: python3 physics_decoder_experiment.py
Stage 2 — Benchmark vs. tuned baseline (arXiv:2607.06287)
Significantly WORSE than the tuned baseline (+154.0%, p=0.0083, wins 0 / 8 paired seeds).
- Benchmark
- Poisson_2d_sparse_boundary
- Model
- fourier_coordinate_mlp
- Paired seeds
- 8
- Baseline mean
- 0.256
- Idea mean
- 0.6502
- Effect (Δ)
- 0.3942 (+154.0%; negative = idea better)
- Wins
- 0 / 8 paired seeds
- p-value
- 0.0083 (permutation test, 20 000 shuffles)
- Smallest detectable effect
- n/a
- Mechanism
- Confirmed ✓
- Practical effect
- Hurts
- Baseline tuning
- swept over 3 configs
The advertised shared bench package and README were absent from the filesystem, so the mandated custom Poisson track and equivalent paired protocol were run locally. No FNO or DeepONet was tested; the intervention was implemented as a training PDE residual rather than per-sample inference-time decoder optimization.
How to run: python3 run_poisson_bench.py
Verdict computed by deterministic test code from paired-seed statistics — not by the language model.
Verification history
Stage 2 — Benchmark Failed ✗
Agent confidence: 9/10
I selected the dynamics track because this physics-informed decoder is closest to the built-in controlled pendulum system. The standard rnn_small baseline achieved mean MSE 0.003223, while the inference-time physics rollout blend achieved 0.012347 across the same 8 seeds, losing on all seeds; the paired permutation p-value was 0.0081. Thus the idealized contraction mechanism holds algebraically, but the fitted rollout mismatch overwhelms the benefit and the idea does not transfer as a useful benchmark win.
- Baseline
- Tuned rnn_small: mean MSE 0.0032230460, std 0.0020170552, lr=0.003, weight_decay=0.0
- Idea
- Physics rollout blend alpha=0.4: mean MSE 0.0123465359, std 0.0012228652; 0/8 wins; paired delta +0.0091234899; p=0.0081
The benchmark uses a pendulum rollout proxy rather than the proposed 2D Poisson coordinate-MLP decoder with automatic second derivatives and online optimization. It reports test MSE only; PDE residual, boundary violation, inference optimization wall-clock, and a genuine differentiable reconstruction loop were not tested. The mechanism signature is an idealized blend identity, not an independent transfer of the Poisson eigenvalue law.
How to run: python3 bench_physics_decoder.py
Stage 2 — Benchmark latest Failed ✗
Agent confidence: 9/10
The matched 2D Poisson benchmark used a tuned data-only Fourier coordinate-MLP baseline and an autodiff PDE-residual arm over 8 paired seeds. The physics arm was worse on the primary relative L2 metric, boundary error, and training time, so the idea did not achieve a win. The analytical contraction mechanism was confirmed, but it did not transfer to this neural-network training setting.
- Baseline
- Mean relative L2 0.2559653; boundary RMSE 0.0702808; mean training time 0.2245203 seconds; best lr=0.01, epochs=80.
- Comparison
- Idea minus baseline = +0.3941881 relative-L2 points; idea wins 0/8 pairs; permutation p=0.00834958.
- Idea
- Best mean relative L2 0.6501534; boundary RMSE 0.0988070; mean training time 1.6559514 seconds; best lr=0.01, epochs=80, mu=0.0001.
Benchmark result
Significantly WORSE than the tuned baseline (+154.0%, p=0.0083, wins 0 / 8 paired seeds).
- Benchmark
- Poisson_2d_sparse_boundary
- Model
- fourier_coordinate_mlp
- Paired seeds
- 8
- Baseline mean
- 0.256
- Idea mean
- 0.6502
- Effect (Δ)
- 0.3942 (+154.0%; negative = idea better)
- Wins
- 0 / 8 paired seeds
- p-value
- 0.0083 (permutation test, 20 000 shuffles)
- Smallest detectable effect
- n/a
- Mechanism
- Confirmed ✓
- Practical effect
- Hurts
- Baseline tuning
- swept over 3 configs
The advertised shared bench package and README were absent from the filesystem, so the mandated custom Poisson track and equivalent paired protocol were run locally. No FNO or DeepONet was tested; the intervention was implemented as a training PDE residual rather than per-sample inference-time decoder optimization.
How to run: python3 run_poisson_bench.py
Artifacts
- README.md 2.0 KB View
- bench_physics_decoder.py 3.9 KB View Raw
- bench_report.json 4.7 KB View Raw
- custom_poisson_track.py 2.6 KB View Raw
- physics_decoder_experiment.py 4.5 KB View Raw
- report.md 1.6 KB View
- report_bench_2026-08-30T121406.md 3.9 KB View
- report_bench_2026-08-30T125215.md 5.5 KB View
- results.json 3.6 KB View Raw
- run_poisson_bench.py 5.2 KB View Raw