Distribution-Level Active Learning for Probabilistic Characterization of Pore Properties in Laser Powder Bed Fusion
Porosity governs the performance of laser powder bed fusion (LPBF) parts, yet it is usually characterized by scalar summary statistics that discard the variability and multimodality of real pore populations. We present a probabilistic framework that predicts full distributions of pore sphericity and maximum Feret diameter as continuous functions of laser power and scan velocity. The pore population of each 316L specimen, measured by micro-computed tomography, is represented as a four-component lognormal mixture, and multitask Gaussian processes learn how the mixture parameters vary with process settings. The surrogate captures distribution shifts across keyhole, optimal, and lack-of-fusion regimes, confirmed by leave-one-out cross-validation. We also introduce an active-learning acquisition function that targets epistemic uncertainty in the predicted density. In synthetic campaigns, it achieves a given accuracy in fewer experiments than grid and random sampling do for the Wasserstein-1 distance, Kullback–Leibler divergence, and log-predictive density. This offers a general approach to probabilistic pore characterization and efficient experimental design.
Authors
- Maher B. Alghalayini (ORCID: https://orcid.org/0000-0002-4071-0272)
- Nadia Kouraytem (ORCID: https://orcid.org/0000-0002-7183-2774)
- Marcus Noack
- Tasrif Ul Anwar
Institutions
- Utah State University (US)
- Lawrence Berkeley National Laboratory (US)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-20
- DOI
- https://doi.org/10.21203/rs.3.rs-11790014/v1
- Primary Topic
- Additive Manufacturing Materials and Processes
- Type
- preprint