Migrating from the original scripts

The original repository was a folder of scripts you ran from inside folder_cfd/ (run.sh -> mpi_tot.py). The code is the same code, reorganized; the old names still work.

Original	Now
folder_cfd/csdl_dafoam.py: instantiateDAFoam	csdl_dafoam.solver.instantiate_dafoam (instantiateDAFoam kept as an alias)
csdl_dafoam.py: DAFoamSolver, DAFoamFunctions	csdl_dafoam.solver
csdl_dafoam.py: compute_dafoam_input_variables	csdl_dafoam.inputs
folder_cfd/csdl_idwarp.py: DAFoamMeshWarper	csdl_dafoam.mesh_warp
folder_cfd/standard_atmosphere_model.py	csdl_dafoam.atmosphere
folder_cfd/bwb_helper_functions.py (setup_geometry, pickle helpers, gather_array_to_rank0)	csdl_dafoam.geometry (setup_geometry kept as an alias), csdl_dafoam.mpi_utils
folder_cfd/mpi_tot.py	csdl_dafoam.airfoil.build_airfoil_model plus examples/airfoil_optimization.py --case naca0012
run.sh	mpirun -np 4 python examples/airfoil_optimization.py --case naca0012
folder_cfd/openfoam_naca0012/, folder_geo/.../airfoil_transonic_unitspan_2.stp	bundled data, csdl_dafoam.cases
Docker image cfdkang/csdl_dafoam	conda environment (Installation)

Behavior changes

• Volume-mesh deformation defaults to IDWarp-JAX (warper="jax") instead of the Fortran IDWarp; the optimizer defaults to modOpt’s OpenSQP instead of PySLSQP. warper="idwarp" and algorithm="PySLSQP" reproduce the original choices.

• Failed mesh check or primal solve raises DAFoamError (originally: skipped / ignored). See How it works.

• DAFoam calls run in the case directory (originally only the adjoint solve did).

• Geometry and projection caches go to cache_dir, keyed by a digest of the STEP file and the surface mesh; the committed stored_files/*.pickle and __pycache__ files are gone, and a cache that no longer unpickles is regenerated instead of crashing.

• lsdo_geo imports use the current top-level names (lg.SectionalParameterization, …), falling back to the deprecated VolumeSectionalParameterization names on older versions.

• Projection is serial by default (projection_options={"num_workers": N} to parallelize).

• The OpenFOAM case no longer ships a stale blockMeshDict (it described a different geometry) or controlDict~.

• gather_array_to_rank0 uses pickle-based comm.gather instead of Gatherv with explicit MPI datatypes: same result, no mpi4py import in the module, testable without MPI.

• The debugging switches of mpi_tot.py (faulthandler, PETSC_OPTIONS=-malloc_debug, interactive vedo plots) are not carried over.

Version note

The original ran on a Python 3.9 Docker image (dafoam/opt-packages, OpenFOAM v1812-era DAFoam and csdl_alpha’s dev branch). The packaged stack targets DAFoam v5.1.1 on OpenFOAM v2506 and current csdl_alpha main. The DAFoam methods used are identical in v4.0.4 and v5.1.1 and csdl_alpha main now contains the MPI region operations, but the bundled case files were written for v1812 and have not yet been run on v2506 (Verification status).