Bio:
Dr. Cadmus Yuan is currently a Professor in the Department of Mechanical and Computer-Aided Engineering at Feng Chia University. Dr. Yuan graduated from National Tsing Hua University in Hsinchu, obtaining his Ph.D. in Power Mechanics in 2005, with a major in Solid Mechanics. After graduation, he conducted postdoctoral research at the Technology University of Delft in the Netherlands. He is focusing on applying AI method and Finite Element Method (FEM) to semiconductor packaging design, reliability issues, and smart manufacturing technologies.
Abstract:
With the advent of cutting-edge technologies such as artificial intelligence (AI), autonomous driving, and heterogeneous integration, electronic packaging has evolved toward more complex and miniaturized architectures—such as chiplet-based systems and high-density interconnects. These advances bring significant challenges to mechanical reliability, particularly concerning solder joint fatigue in plastic ball grid array (PBGA) packages under thermal and mechanical loads.
To address such reliability concerns, finite element modeling (FEM) has become a widely adopted tool for virtual prototyping. While FEM offers high-fidelity simulations, it remains computationally intensive and requires domain expertise for material calibration, boundary condition settings, and result interpretation. These constraints render FEM impractical for early-stage design iterations that demand rapid evaluation over vast design spaces.
This invited talk introduces a hybrid framework that integrates AI-based surrogate models with conventional FEM simulation workflows. Rather than replacing FEM, our framework leverages the complementary strengths of both approaches: AI models for rapid inference and exploration, and FEM for high-accuracy validation and critical case analysis. This co-existence strategy enables scalable, cost-effective reliability design in industrial settings.
A deep neural network (DNN) architecture is proposed as the core surrogate engine. The model is trained on deterministic FEM-generated datasets and designed to capture multi-dimensional, nonlinear design-to-response mappings. Once trained, the surrogate model enables fast approximations, optimization via gradient descent (GD) or particle swarm optimization (PSO), and large-scale Monte Carlo simulations for sensitivity analysis and uncertainty quantification. The resulting insights can be used to filter and prioritize design candidates before passing them to FEM for detailed verification.
This talk will highlight:
The motivation and architecture of the AI-FEM co-design framework.
The methods to train the AI model with new training indicators
The application of the surrogate model in reliability optimization and uncertainty assessment within the AI-FEM loop.
By embedding intelligence early in the design process, this framework bridges physics-based modeling and data-driven inference, enabling more efficient and informed design decisions for next-generation electronic packaging.