TCPC25_03 Non-Newtonian LBM Setup

This repository combines the working TCPC phantom LBM setup with the older Carreau-Yasuda non-Newtonian LBM implementation. It is configured for the #04 PE simulace regimes v1, v2, v3, v4a, and v4b, each runnable as:

• newtonian
• sx (saccharose/SX constants from the historical non-Newtonian code)
• gx (glycerol/GX constants from the historical non-Newtonian code)

Generated geometry.vtk, solver binaries, and simulation results are intentionally not tracked.

Prerequisites

Python:

python3 -m pip install numpy scipy trimesh vtk pyyaml

Solver machine:

• CUDA compiler (nvcc)
• MPI compiler wrapper (mpicxx)
• VTK C++ development headers/libraries
• TNL checkout compatible with the historical solver

The build uses TNL_DIR to find TNL. Example:

scripts/build_solver.sh TNL_DIR=/home/bures/geraldine/Eichler/2026/2026_02_29-nonNew/Glycerol-nonNew/tnl_submodule

If VTK is not discoverable through pkg-config, pass VTK_CONFIG:

scripts/build_solver.sh \
  TNL_DIR=/path/to/tnl_submodule \
  VTK_CONFIG="-I/usr/include/vtk -lvtkCommonCore -lvtkIOLegacy -lvtkCommonDataModel -lvtkIOXML -lvtksys"

Geometry

Generate the runtime geometry from the tracked STL:

python3 scripts/generate_geometry.py

This writes:

• runtime/geometry.vtk: solver input, with wall labels
• runtime/test.vtk: pre-PE label check
• runtime/TCPC25_03_tapered.stl: tapered/capped intermediate STL

The default geometry reproduces the historical PE setup: 25 mm outlet extensions and 4 mm outlet diameter.

The generated VTK file is used as the wall mask. To match the historical working TCPC solver, the simulation imposes open boundary planes itself: x-min is IVC, x-max is SVC, y-min is outlet label 5, and y-max is outlet label 6. Wall cells from the VTK geometry override those plane labels.

Build

scripts/build_solver.sh TNL_DIR=/path/to/tnl_submodule

The solver executable is:

solver/sim_tcpc/sim_tcpc

The build still needs TNL headers for the lattice arrays, but the unused immersed-boundary TNL sparse-matrix branch is disabled by default. Re-enable it only for IBM/Lagrange experiments with use_TNL_LAGRANGE=yes.

Run Cases

Print all configured commands:

python3 scripts/run_all_cases.py --dry-run

Run one case:

python3 scripts/run_case.py --regime v1 --material sx

For pilot or production runs, override timing/output directly on the command line instead of editing YAML:

python3 scripts/run_case.py --regime v1 --material newtonian --final-time 1 --vtk-period 1
python3 scripts/run_all_cases.py --final-time 10 --vtk-period 1 --print-period 0.1

The solver writes results under results_<case-id>/, for example results_v1_sx/.

The command generated for each case passes the prescribed IVC/SVC flow rates from configs/regimes.yaml directly in ml/s.

Slurm Batch Runs

For production runs, use the Slurm orchestrator from a persistent screen or tmux session. It writes one job.sbatch and manifest.json per case, submits the jobs, polls Slurm until all jobs finish, then computes flows.csv for each completed result directory and a combined flow_summary.csv under runtime/slurm/run-*.

Dry-run the planned submissions:

python3 scripts/slurm_run_cases.py \
  --dry-run \
  --case-prefix prod_batch_ \
  --case-suffix _t10 \
  --final-time 10 \
  --vtk-period 1 \
  --print-period 0.1

Submit all configured regimes/materials and wait for postprocessing:

python3 scripts/slurm_run_cases.py \
  --case-prefix prod_batch_ \
  --case-suffix _t10 \
  --final-time 10 \
  --vtk-period 1 \
  --print-period 0.1 \
  2>&1 | tee runtime/slurm_driver_$(date +%Y%m%d_%H%M%S).log

Defaults request --partition=gp, one GPU, 8 CPU threads, 32 GB RAM, and 12 hours walltime per case. Override these with flags such as --time 24:00:00, --mem 64G, --gres gpu:rtx5080:1, or --partition PARTITION_NAME.

The generated jobs export OMPI_MCA_accelerator=cuda so Open MPI does not abort on nodes where both ROCm and CUDA accelerator plugins are present. Additional job environment variables can be supplied as --env NAME=VALUE.

If the screen session is interrupted but Slurm jobs keep running, resume polling and harvesting with:

python3 scripts/slurm_run_cases.py --resume runtime/slurm/run-YYYYMMDD-HHMMSS

Outputs

VTK output includes:

• wall
• velocity
• mean_velocity
• lbm_rho
• pressure_pa
• gamma_dot
• nu_phys_m2_s
• mu_phys_pa_s
• turbulence_intensity
• rms_velocity
• strain_diag_1_s
• strain_shear_1_s
• wss_proxy_pa

wss_proxy_pa is an approximate voxel-gradient wall-shear proxy on fluid cells adjacent to wall cells. It is not an exact stress projected onto the original STL surface.

Flow Split Postprocessing

Compute inlet/outlet flows and LPA/RPA split from VTK files:

python3 postprocess/compute_flows.py 'results_v1_sx/vtk3D/data_*.vtk' --output results_v1_sx/flows.csv

Default label names are:

• 3:IVC
• 4:SVC
• 5:RPA
• 6:LPA

Override if visual inspection of the generated geometry shows the pulmonary labels are swapped:

python3 postprocess/compute_flows.py 'results_v1_sx/vtk3D/data_*.vtk' \
  --labels 3:IVC,4:SVC,5:LPA,6:RPA \
  --output results_v1_sx/flows.csv

Source Directories Used

• Newtonian TCPC geometry/flow setup: 2026_02_20-simulationSetup
• Non-Newtonian Carreau-Yasuda LBM implementation: 2026_02_29-nonNew
• Historical flow postprocessing reference: 2026_03_18-zpracFlows