SparseThread-AM: What Evidence Can Be Discarded from a Transferable Laser Powder Bed Fusion Quality Record?

Laser powder bed fusion (LPBF) produces increasingly rich manufacturing records, but retaining every available variable does not necessarily produce a more transferable machine-learning model. We study which information is actually required for cross-study high-density prediction using a public compilation of 1,579 LPBF relative-density measurements. Six progressively richer manufacturing records are evaluated under source-held-out validation, beginning with core process variables and sequentially adding material and powder information, geometry, printer identity, measurement provenance, and physics-derived descriptors. Core process variables alone achieve an area under the receiver operating characteristic curve (AUROC) of 0.569 and average precision (AP) of 0.365, while the complete record reaches 0.659 AUROC and 0.500 AP. Importantly, feature additions are not uniformly beneficial. Measurement provenance produces a late-stage improvement of 0.051 AUROC and 0.077 AP, showing that information describing how evidence was generated can be as important as additional process variables. Under a strict prespecified retention criterion, the complete record is required, while a relaxed criterion permits a substantially smaller representation. These results suggest that manufacturing data reduction should preserve transferable evidence rather than simply minimize feature count.