A neural-network approach for semiconductor wafer post-sawing inspection

Abstract

Semiconductor wafer post-sawing requires full inspection to assure defect-free outgoing dies. A defect problem is usually identified through visual judgment by the aid of a scanning electron microscope. By this means, potential misjudgment may be introduced into the inspection process due to human fatigue. In addition, the full inspection process can incur significant personnel costs. This research proposed a neural-network approach for semiconductor wafer post-sawing inspection. Three types of neural networks: backpropagation, radial basis function network, and learning vector quantization, were proposed and tested. The inspection time by the proposed approach was less than one second per die, which is efficient enough for a practical application purpose. The pros and cons for the proposed methodology in comparison with two other inspection methods, visual inspection and feature extraction inspection, are discussed. Empirical results showed promise for the proposed approach to solve real-world applications. Finally, we proposed a neural-network-based automatic inspection system framework as future research opportunities.