Proxy-Label Reconstruction and Baseline Testing in LPBF Thermography Machine Learning

Image-based machine learning is now common in additive manufacturing moni- toring, but high accuracy is hard to interpret when the label is computed from the same image given to the model. This paper studies that problem with the public NIST AM-Bench 2022 Time-Above-Melt (TAM) thermography data for LPBF of IN718. We define a two-class Low-Coverage/High-Coverage task from the fraction of nonzero TAM pixels and build 936 layer images from three repeat builds. The primary split keeps builds separate: B6 for training, B7 for validation, and B8 for testing. Deep image classifiers are compared with shallow scalar-feature mod- els and with a deterministic reconstruction rule that applies the original coverage threshold directly. On B8, Swin-Small at 512 pixels and the deterministic rule each make 0 errors (macro-F1 1.000), while a linear SVM using five descriptors, two of which jointly encode coverage, makes 1 error (macro-F1 0.997). In a five-seed cyclic repeat-build analysis, Swin-Small was repeated with five random seeds for each train/validation/test assignment; the three fold mean macro-F1 values were 0.997 ± 0.003, 0.996 ± 0.004, and 0.991 ± 0.005. The result is simple but important for reporting: a build-wise split can remove same-build image leakage, yet it does not make a proxy label independent of the input image. For this task, near-perfect accuracy means recovery of a visible coverage statistic, not defect prediction or part-quality assessment.