Marcin Grabski

2024-11

Built Abaqus simulation database for ML surrogate model training — 1200 analysis results with geometric parameters, material properties, load levels, and output quantities (max stress, displacement, life).

Challenge: HDF5 file format was chosen for the database to handle large numerical arrays efficiently. However, concurrent write access from multiple HPC jobs required careful file locking. Switched to individual per-job CSV files merged in a post-processing step.

Learned: For large simulation databases generated by parallel HPC jobs, write-then-merge is safer than concurrent shared file access. Each job writes to its own output file; a final merge step combines them. Simpler and more robust than distributed file locking.

2024-05

Integration of Abaqus with Dakota optimization framework for shape optimization of disk bore geometry — minimize weight subject to max von Mises stress < yield / safety factor.

Challenge: Dakota's direct Abaqus interface uses file-based communication — Dakota writes parameter file, Abaqus reads and runs, Dakota reads result file. Race conditions occurred on shared NFS filesystem when multiple evaluations ran simultaneously.

Learned: File-based optimizer-simulator interfaces on shared filesystems need explicit job isolation: unique working directories per evaluation, explicit synchronization before result reading, and cleanup of temp files to avoid disk overflow.

2023-12

Automated report generation from Abaqus ODB files — extracts key results (max stress location and value, deformation, contact status), generates plots using matplotlib, compiles into PDF via ReportLab.

Challenge: Consistent extraction of 'max stress location' is harder than it sounds — results must be filtered to avoid singularities at boundary condition application points and at contact interfaces. Added automatic singularity detection based on local mesh refinement ratio.

Learned: Automated FEM result reporting requires singularity handling. Stress concentrations at constraints are numerical artifacts. A simple filter: if an element's stress is >3x its neighbors' average stress, flag it for manual review rather than reporting it as the maximum.

2023-08

Parametric cooling channel geometry generation for HPT blade — Python script creates Abaqus model geometry, mesh, and boundary conditions for 35 cooling hole diameter and pitch combinations.

Challenge: Parametric geometry generation in Abaqus/CAE Python API is fragile — small changes in part geometry can cause feature-based operations (cuts, fillets) to fail silently and produce wrong geometry. Validated each geometry against a reference mesh by checking element count and volume.

Learned: Parametric Abaqus models built with CAE Python API need geometric validation checks after each generation. Relying on job completion status alone is insufficient — a wrong geometry can produce a valid analysis with wrong results.

2023-03

Built Python scripting framework for automated Abaqus job submission and ODB post-processing — reads parameters from CSV, generates input files, submits jobs, extracts stress/strain results from ODB files, outputs to structured CSV.

Challenge: Abaqus Python API uses a non-standard environment — many common Python packages (pandas, numpy) are not available in Abaqus's embedded Python. Had to run ODB opening inside Abaqus Python and pass extracted data to a separate standard Python process via JSON files.

Learned: Abaqus Python scripting requires awareness that the Abaqus Python interpreter is a different environment from system Python. Separate the ODB reading (inside Abaqus Python) from data processing (system Python) for maximum flexibility.