Adebanji Adelowo

PINO Darcy

Physics-loss weighting in a Darcy-flow neural operator.

Status: Ongoing research. Results below are from the specified reduced-budget experiment, not the upstream full-budget configuration.

Experimental setup

Darcy_241 benchmark, 15 epochs, three training seeds, physics weight 0.1 and data-only control.

Results achieved

Recorded results and study scope
MetricResult
ResidualReduced under physics weighting in all three seeds
Prediction-error difference−0.0004 mean (seed interval −0.0070 to +0.0061)
Full-budget CUDA studyPending

Experimental figures

Comparison of prediction error and PDE-residual trade-offs across physics-loss settings
Prediction error and PDE-residual trade-offs under physics-loss weighting · Full figure ↗
Per-seed comparison of reduced-budget data-only and physics-weighted experiments
Data-only versus physics-weighted PINO models across three training seeds · Full figure ↗

Figures are drawn directly from the project's public GitHub repository.

Interpretation and scope

A lower PDE residual did not establish a consistent prediction-accuracy improvement across the three seeds.

Next steps

Verify the corrected forcing/residual and run the upstream 50-epoch CUDA comparison.

For reproducibility, consult the repository's configuration files, scripts, saved results and scientific audit.

Repository, methods and results ↗ ← All PhysicsNeMo projects