Wojciech Kowalczyk

2025-02

Supported AI team in generating labeled FEM result dataset for surrogate model training — ran 150 parametric ANSYS analyses with varied geometry and load parameters, extracted max stress and displacement as target variables.

Challenge: Automating extraction of maximum stress with spatial location (not just the scalar value) from ANSYS result files required scripting in ANSYS APDL. The built-in ANSYS result reader in Python (ansys-mapdl-core) was more reliable than parsing .rst files directly.

Learned: Simulation data for ML surrogate models must include not just output values but also the input parameter vector and mesh quality metrics. Models trained on poorly meshed outlier cases perform worse on new geometry.

2024-09

Prestress modal analysis of rotating HPT disk — centrifugal stiffening effect on blade natural frequencies. Static analysis with rotational velocity, then modal on prestressed structure.

Challenge: Forgotten PSTRES ON command in the restart caused the modal to ignore prestress. Results looked plausible but were wrong — frequencies were 8% too low. Caught during correlation with spin rig test data.

Learned: Prestress modal analysis has several easy-to-miss setup steps. Building a checklist (NLGEOM ON, PSTRES ON, correct restart point, correct load state) prevents silent errors that produce plausible-looking wrong results.

2024-04

Submodeling workflow for stress concentration around cooling holes in HPT blade — global coarse model provides displacement BCs for local refined submodel around each hole.

Challenge: Submodel cut boundary was placed too close to the stress concentration — violated St. Venant's principle. Results were sensitive to cut boundary location. Moved boundary 3x the hole diameter away and results stabilized.

Learned: Submodel cut boundary placement requires engineering judgment based on Saint-Venant's principle. A parametric study varying the cut boundary location is the only reliable way to confirm independence.

2023-11

Nonlinear buckling analysis of thin-walled titanium bracket — geometric nonlinearity (NLGEOM ON), initial imperfection seeded from first eigenmode shape at 1% of characteristic dimension, arc-length method.

Challenge: Arc-length method (Riks) was losing the equilibrium path at the limit point. Had to manually reduce the arc-length increment and increase the maximum number of increments. Post-buckling behavior showed snap-through — physically meaningful but numerically tricky.

Learned: Nonlinear buckling with snap-through requires Riks or displacement-controlled loading. Force-controlled loading cannot pass the limit point. Always seed a realistic imperfection — perfect geometry can give non-conservative bifurcation loads.

2023-07

Bolted flange contact analysis for HPT casing joint — 16-bolt pattern, friction contact (mu=0.15), bolt preload via pretension element, operating load applied as second load step.

Challenge: Contact status oscillation (open/closed/sliding cycling) preventing convergence. Augmented Lagrange formulation was less stable than Pure Penalty for this configuration. Switched to Pure Penalty with reduced normal stiffness factor (FKN=0.1).

Learned: Contact formulation choice depends on the physical problem. Augmented Lagrange is generally better for large deformation; Pure Penalty converges more reliably for stiff contacts with small relative sliding.

2023-02

HPT disk nonlinear thermal-structural analysis — temperature-dependent material properties (elastic modulus, CTE, yield strength as function of T). Plasticity model: bilinear isotropic hardening for Inconel 718.

Challenge: Convergence issues with temperature-dependent material near the yield knee — solver oscillated between elastic and plastic states. Resolved by activating line search and tightening the force convergence criterion from 0.5% to 0.1%.

Learned: For thermal-structural analyses with plasticity, the load step definition matters enormously. Ramping the thermal load smoothly over many substeps is more important than tight convergence tolerances.