AI-Driven Defect Classification in Advanced Packaging: Lessons from Production Deployments
Advanced packaging has become the critical layer enabling AI compute — from co-packaged optics linking AI accelerators inside data centers, to high-density CIS modules powering AI vision in autonomous systems, to fan-out and 2.5D/3D integration of next-generation devices. Yet as packaging complexity rises, rule-based inspection increasingly fails to keep pace with new defect modes: RDL opens and bridging, bump deformation, TSV voids, die tilt, warpage, and interfacial defects in hybrid bonding. Critical defects escape into high-cost downstream steps, while false rejects from over-tuned recipes erode throughput and operator trust. Inspection is no longer the bottleneck; classification is.
This session draws on production experience across advanced-packaging segments in Taiwan and Singapore: optical interconnects for AI data-center infrastructure, CIS packaging for AI vision and autonomous-system sensors, fan-out packaging, and 2.5D/3D integration including co-packaged optics. Each represents a distinct defect environment, labeling challenge, and integration constraint.
The talk will cover:
- Defect taxonomy and labeling strategies across optical-interconnect, fan-out, sensor packaging, and 2.5D/3D process flows
- Model design choices for the low-data, high-variability conditions characteristic of new packaging nodes
- Deployment patterns for integrating AI-driven classification into existing inspection toolsets and MES/SECS-GEM workflows
- Cross-deployment results: reductions in false rejects, recovered operator capacity, and earlier detection of process drift
Attendees will leave with a practical framework for evaluating where AI-driven classification fits within their own advanced-packaging inspection strategy — and a clear-eyed view of what production deployment requires beyond model accuracy.
Key Technologies Covered
- AI-Driven Defect Classification for Heterogeneous Integration
- Hybrid bonding interfacial defect detection
- Self-supervised learning model design for low-data, high-mix packaging environments
- Active learning and human-in-the-loop labeling for new defect classes
- Transfer learning across packaging segments and process nodes
- Multi-modal inspection data fusion (optical, SEM, X-ray)
- AI integration with inspection toolsets and MES/SECS-GEM workflows
- Anomaly detection for early identification of process drift
