Capturing mold qualification sensitivities for adaptive startup tuning in injection molding

Injection-molding changeovers trigger production startups where parameters are re-tuned to recover the defect-acceptance condition, generating scrap when qualification settings are reused without adaptation. In this work, qualification is reframed as knowledge acquisition, capturing defect-parameter sensitivities for decision support. Two assistants trained on the same dataset are compared: a predictive assistant with delta-anchored tuning and a retrieval-grounded assistant reusing qualification evidence for constrained updates. Vision inspection supplies feedback. Validated on a socket cover across hydraulic and electric machines and three viscosities, the proposed assistance reduced median runs-to-quality by 88% and mean startup scrap by 94% relative to the controlled baseline.