
AI Reduces Advanced Semiconductor Packaging Design Time by 1,000 Times
The AI Semiconductor Process Technology and Design Research Team from the Department of AI Semiconductor Engineering at Korea University Sejong Campus presented its research achievements on AI-based signal integrity optimization for organic RDL interposers at the international semiconductor packaging conference, the 2026 IEEE 76th Electronic Components and Technology Conference (ECTC).
The study, titled “AI-Driven Inverse Design and Vision Fusion Framework for Automated Signal Integrity Optimization of Organic RDL Interposers,” was conducted by researchers Jo Minseon (first author), Lee Taeheon (co-author), Yoo Jiho (co-author), Kim Dayoung (co-author), and Professor Jeong Seongyeop (corresponding author).
The paper was selected for an oral presentation and was published in the IEEE Conference Proceedings.
From May 26 to 30, The 76th ECTC was held at the JW Marriott & The Ritz-Carlton Grande Lakes in Orlando, Florida, USA.
ECTC is one of the leading international conferences where the latest research achievements and industry trends in semiconductor packaging, electronic components, and microelectronic systems are shared. It is widely recognized as a major forum for technical exchange among academic and industry experts worldwide.
With the rapid growth in demand for Artificial Intelligence (AI), High-Performance Computing (HPC), and High-Bandwidth Memory (HBM), advanced 2.5D and 3D packaging technologies that densely integrate multiple chips into a single package have become increasingly important. In particular, RDL interposers serve as critical structures that provide stable electrical signal connections between chips and are regarded as key components determining the performance and reliability of next-generation semiconductors.
However, in actual manufacturing processes, copper interconnects are often formed with trapezoidal rather than ideal rectangular cross-sections, making it difficult for conventional equation-based models to accurately predict signal characteristics. Although high-precision electromagnetic simulations offer excellent accuracy, they require substantial time even to validate a single design point, resulting in significant time and cost expenditures during iterative design and optimization processes.
To address these challenges, the research team proposed an AI-based automated design framework. The framework integrates three key components into a single closed-loop system: a surrogate model that predicts impedance from wiring structures, an inverse design model that identifies optimal structures based on target performance requirements, and a vision-based classifier (CNN) that analyzes eye-diagram images to determine signal quality compliance.
Using this framework, the team demonstrated that design time could be significantly reduced while maintaining accuracy comparable to conventional simulation methods.
The results showed prediction accuracies of R² = 0.997 for single-ended signal structures and R² = 0.994 for differential signal structures. Furthermore, the framework demonstrated the potential to reduce design time by more than 1,000 times compared with conventional approaches.
The researchers also provided data-driven evidence that the impedance-matching-centered design procedures widely used in the industry may not be sufficient to guarantee actual signal integrity. Even when impedance values fall within target ranges, issues may still arise in practical signal quality metrics such as eye diagrams. The AI-based vision model was shown to effectively identify and evaluate such issues.
This research holds both academic and industrial significance because it combines AI-driven inverse design and vision-based signal integrity verification into a unified automated framework. In particular, it provides a foundation for the rapid and precise design automation technologies required for next-generation semiconductor packages, including HBM, AI accelerators, and high-performance server semiconductors.
First author Jo Minseon stated, “Rather than relying on conventional equation-based analytical models or repeatedly performing simulations, we were able to build more accurate and efficient design models by utilizing the function approximation capabilities of artificial neural networks and deep learning. In particular, we demonstrated that data-driven AI can accurately and efficiently predict and verify signal integrity mismatches that are difficult to detect through impedance matching alone. It was especially meaningful to present these achievements at ECTC and engage with researchers from around the world.”
Corresponding author Professor Jeong also commented, “I am pleased that we achieved meaningful results through interdisciplinary research combining artificial intelligence and semiconductor technologies.”
This research was supported by the Korea Evaluation Institute of Industrial Technology (KEIT; RS-2024-00417909, RS-2025-25452526), KCHIPS (25081-15FC), and Korea University. Additional support was provided through EDA tools from the Integrated Circuit Design Education Center (IDEC) and technical assistance from Synopsys Korea.
Office of External Relations, Korea University Sejong Campus