Mechanism-Driven Deep Learning for Predicting the Impact of Hard Impurity Particles on Direct Wafer Bonding Performance
Abstract
Direct wafer bonding is a foundational technology for fabricating micro-electro-mechanical systems (MEMS) and achieving high-density multifunctional chip integration. However, the presence of environmental impurity particles at the bonding interface remains a critical challenge. While traditional evaluation methods rely on finite element analysis (FEA) and analytical mechanical models, they face high computational costs for complex, non-uniform impurity distributions. This paper proposes a mechanism-driven machine learning framework based on Physics-Informed Neural Networks (PINNs) to efficiently predict the bonding quality of wafers containing hard impurity particles. The PINN algorithm embeds fundamental mechanical governing equations — relating to strain energy, wafer deflection, and adhesive contact mechanics—directly into the loss function of the deep learning architecture. Input features encompass external normal pressure, wafer geometric parameters, and diverse impurity distribution patterns (Cluster, Complex, Face, Line). Results demonstrate that wafer curvature, thickness, and impurity distribution jointly dictate the strain energy under specific pressures, with the Face distribution exerting the most detrimental influence. The interface porosity was controlled below 2.3%, while predicted bonding strength reached 18.5 MPa. The neural network achieves a computational speedup of nearly three orders of magnitude (approximately 600 ×) compared to traditional FEA while maintaining equivalent physical accuracy.