Skip to content

Self-Training ML-Assisted CFD Workflow

Choose a tag to compare

@pauloacs pauloacs released this 18 Oct 09:36
· 7 commits to main since this release
bdc5cca

This release introduces an integrated machine-learning workflow for initializing OpenFOAM simulations with learned pressure and optional velocity estimates, followed by normal PIMPLE correction.

Highlights

  • New DLpimpleFoam_auto workflow for sampling simulation data, training surrogate models, and using predictions during CFD runs.
  • Two surrogate backends: the existing gridded model, now named pressure_SM_gridded_shift, and the new mesh-graph package, pressure_SM_mesh_graph.
  • Expanded prediction options: pressure increments can use dp or ddp targets; joint models can predict velocity increments using dU or ddU.
  • Improved mesh transfer with volume-preserving, slope-limited prolongation and configurable native-baseline composition.
  • More flexible state injection, including configurable pressure reconstruction, direct handoff to PIMPLE, and optional timestep-scaled velocity extrapolation.
  • New DLpimpleFoam_oracle solver for diagnostic comparisons of pressure and velocity integration strategies.
  • Runnable tutorials for gridded and graph pressure-only workflows, graph joint prediction, and oracle cases.