Forty Examples Can Change the Deployment Regime: Few-Shot Cross-Alloy Adaptation for High-Density Laser Powder Bed Fusion Screening

Machine-learning models for laser powder bed fusion (LPBF) are often trained by pooling observations from multiple alloys, but deployment may require prediction on an alloy for which little labeled quality data are available. We study a practical calibration question using a public compilation of 1,579 LPBF relative-density measurements spanning six commercial alloys. The binary task screens whether relative density reaches 99%. In a leave-one-alloy-out few-shot protocol, all non-target alloys provide the base training set and only a small labeled calibration set is added from the target alloy. Averaged across the six target alloys, area under the receiver operating characteristic curve (AUROC) increases from 0.556 with zero target examples to 0.628 with 5, 0.661 with 10, 0.733 with 20, and 0.805 with 40. Forty target examples therefore recover 56.0% of the gap between the zero-shot result and perfect AUROC. The result does not establish material-invariant qualification. It shows that cross-alloy reuse should be evaluated as a sample-efficiency problem: zero-shot transfer can be weak, while a limited target-alloy calibration set can move the model into a substantially more useful discrimination regime.