QualiAI: Uncertainty-Guided Verification Before Qualification in AI-Assisted Metal Additive Manufacturing

Machine-learning systems for metal additive manufacturing can classify process states and predict quality, but qualification-oriented deployment requires a second question: which predictions deserve additional physical verification? We study uncertainty-guided verification using public laser powder bed fusion data with source-held-out evaluation. The overall prediction error rate is 30.21%. When only the 10% most uncertain cases are selected for verification, the selected subset contains 40.76% errors, corresponding to 1.349 times the background error rate. The enrichment remains 1.303 times at a 20% verification budget, 1.243 times at 25%, and 1.162 times at 30%. These results show that uncertainty can help prioritize limited verification resources, although it should not be interpreted as a certification rule. We frame the result as verification before qualification: predictive confidence, physical evidence, provenance, and transfer assumptions should be considered separately before an AI-supported manufacturing decision is accepted.