Speaker
Description
Thermo-mechanical simulation of Laser-Powder Bed Fusion (L-PBF) requires calibration against measured displacement, but the common practice of tuning a single Strain Scaling Factor (SSF) applies one constant to the entire strain field and cannot reproduce the localized displacement caused by non-uniform thermal contraction. We present an active-learning framework for multi-parameter inverse calibration of L-PBF simulation : of 14 thermo-mechanical and material factors, the five most influential on displacement (CTE, Young's modulus, yield strength, SSF, absorptivity) are varied, and a Gaussian-process surrogate of the forward model is trained adaptively. On a cantilever benchmark mapping the five inputs to 251 displacement outputs (ANSYS 2024 R1), each campaign starts from 50 Latin-hypercube samples, adds up to 500 acquisition-selected simulations, and terminates when the five-fold cross-validated RMSE of the surrogate falls below 0.05 on the standardized scale, equivalent to explaining 99.75% of the response variance. Eighteen acquisition variants (expected-improvement sweeps and schedules, probability of improvement, confidence bounds, randomized selection, enlarged pools, random search; 54 campaigns) all failed, plateauing at RMSE 0.0572. The bottleneck is shown to be surrogate specification rather than sampling policy : replacing the isotropic Matérn kernel with an anisotropic kernel that learns one length-scale per input improved accuracy by 25-32% at a fixed data budget and reached the target (RMSE 0.0492 at iteration 384), with all replicates beating every baseline; multi-start verification ruled out optimizer effects. ARD also flagged absorptivity as low-sensitivity. This gain carries a cost : per-iteration hyperparameter refitting grows polynomially with accumulated data, from under 240 s to over 3,000 s by iteration 400. ARD kernels are therefore necessary for L-PBF calibration surrogates, paired with early stopping and reduced refit frequency.
| Speaker Country | Republic of Korea |
|---|---|
| Would you like to publish your paper in the special issue of BHM "Berg- und Hüttenmännische Monatshefte" | No |