Run-to-Run Failure Detection and Diagnosis Using Neural Networks and Dempster–Shafer Theory: An Application to Excimer Laser Ablation

Abstract

The formation of microvias in multilayer substrates is a critical factor in microelectronic packaging manufacturing. Such microstructures can be produced efficiently by excimer laser ablation. Thus, laser ablation systems are evolving to a level where the need to offset high capital equipment investment and lower equipment downtime are imminent. This paper presents a methodology for inline failure detection and diagnosis of the excimer laser ablation process. The methodology employs response data originating directly from the equipment and characterization of microvias formed by the ablation process. Neural network (NN) models are trained and validated based on this data to generate evidential belief for potential sources of deviations in the responses. Dempster-Shafer (D-S) theory is adopted for evidential reasoning. Successful failure detection is achieved in 100% of 19 possible failure scenarios. Moreover, successful failure diagnosis is also achieved with only a single false alarm occurring in the 19 failure scenarios.