Interpretable Machine Learning for Process Parameter Analysis in Arc-Driven Powder Bed Fusion of 316L Stainless Steel

Arc-driven powder bed fusion represents a low-cost alternative to beam-based powder bed systems, yet the morphological stability regimes governing single-track formation and the relative influence of process parameters on regime transitions have not been systematically characterised. Manual visual assessment of track morphology is inherently subjective and cannot objectively quantify the parameter hierarchy governing stability boundaries. This study addresses both limitations through two complementary contributions. A deterministic two-stage image-based framework is developed to automatically classify single-track morphology from top-view images of solidified 316L stainless steel tracks, replacing subjective assessment with a reproducible, intervention-free procedure. A gap-based continuity criterion distinguishes discontinuous from continuous melt paths; for continuous tracks, the coefficient of variation in width (CV (coefficient of variation) < 0.15) further separates geometrically stable from transitional morphologies. Building on the image-derived regime labels, two interpretable classifiers—a depth-limited Decision Tree (DT) and a regularised Logistic Regression (LR)—are fitted using applied current, scanning speed, and electrode-to-powder-bed distance as predictors. The classifiers are employed not for predictive generalisation but to extract standardised coefficients and permutation-based feature importance rankings, yielding a model-agnostic, quantitative explanation of which process parameters govern regime transitions. Stable continuous tracks are obtained only within a restricted parameter window. Permutation importance consistently ranks applied current as the dominant predictor, followed by electrode distance and scanning speed, in agreement with the thermophysical interpretation. Logistic Regression coefficients confirm that reduced stand-off distance is a necessary condition for sufficient arc constriction. Supplementary linear regression models indicate that applied current governs melt pool depth, whereas scanning speed is the primary determinant of width variation. The combined framework establishes a reproducible basis for process parameter hierarchy analysis in arc-driven powder bed systems and provides a foundation for regression-based process optimisation.