Integrating Transfer Learning with Additive Manufacturing Simulation Data for Accelerated Defect Prediction in Multi-Material Fabrication

Abstract

Additive manufacturing (AM) has emerged as a transformative technology across industries, yet defect prediction remains a significant challenge for multi-material fabrication processes. This research introduces a novel transfer learning framework that leverages simulation-generated data to predict real-world defects in multi-material AM components. We develop a domain adaptation architecture that addresses the reality gap between simulation and physical processes through a combination of contrastive learning objectives and physics-informed neural networks. Our approach demonstrates significant improvements in defect prediction accuracy, achieving a 27.3\% reduction in false negative rates compared to traditional machine learning methods. The framework successfully identifies subsurface porosity, layer delamination, and thermal stress-induced defects with 94.8\% precision in polymer-metal composite structures without requiring extensive real-world training data. Performance evaluations across five distinct material combinations reveal robust generalization capabilities. This work presents a paradigm shift in quality assurance for multi-material AM processes by enabling pre-emptive defect detection through computationally efficient transfer learning mechanisms, thereby reducing material waste and post-processing requirements while increasing overall manufacturing reliability.