Five ongoing research studies built around NVIDIA PhysicsNeMo examples. Each study documents the research question, results already obtained and the next planned experiments. Interim results refer to reduced-budget experiments unless stated otherwise.
Research status: Active studies. Results describe the recorded experimental settings and are not full-budget benchmarks. See repositories for the latest run status ↗
Detailed case study: figures, metrics and reproducibility →
1. PINO Darcy
What I am investigating: Physics-loss weighting for Darcy-flow neural operators.
Results achieved: Across three 15-epoch seeds, physics weighting reduced PDE residuals. With a data-consistent residual at weight 0.1, mean prediction-error change was −0.0004 (seed interval −0.0070 to +0.0061).
Next steps: Verify the corrected residual and run the upstream 50-epoch CUDA comparison.
Detailed case study: figures, metrics and reproducibility →
2. FNO Darcy
What I am investigating: Fourier-mode capacity versus accuracy and cost.
Results achieved: Reduced 64×64 multi-seed studies found 8, 12 and 16 Fourier modes broadly similar at the original fixed training budget; a 3×-budget study has also run.
Next steps: Complete CUDA generator checks and 256×256 full-budget and upstream controls.
Detailed case study: figures, metrics and reproducibility →
3. MeshGraphNet Vortex
What I am investigating: Message-passing depth and long-horizon vortex shedding.
Results achieved: Nine reduced-budget models (three depths × three seeds) completed. Greater depth improved one-step prediction; none of 63 assessed rollouts preserved shedding after step 100.
Next steps: Investigate rollout stability and evaluate a longer training budget.
Detailed case study: figures, metrics and reproducibility →
4. Stokes MeshGraphNet
What I am investigating: Physics-informed fine-tuning of a flow surrogate.
Results achieved: On two test meshes, velocity relative L² error decreased 0.188 → 0.109 (41.8%); momentum residual fell 67.5%, divergence residual 89.6%.
Next steps: Complete provenance checks and evaluate meshes 2–9.
Detailed case study: figures, metrics and reproducibility →
5. Lid-driven cavity PINNs
What I am investigating: PINN accuracy across Reynolds numbers.
Results achieved: Reduced studies at Re = 10, 100 and 400 reported mean velocity errors 0.462, 1.072 and 0.946. Reference checks and 52 audit tests passed.
Next steps: Run nine full-budget CUDA experiments (three Reynolds numbers × three seeds).
These are active studies, not claims that the full-budget experiments have been completed.