Test And Reliability Improvement With Defect-Image Classification And Machine-Learning Algorithms In Semiconductor Industry For Automotive Applications

In the semiconductor-manufacturing industry for automotive applications, parts are tested at each manufacturing step to screen likely-to-fail parts, in manufacturing lines or later in the field. In particular, all the parts are optically inspected and electrically tested. However, each of the defects observed at inspection should be submitted to a manual assessment about its potential impact on die reliability, and, because of complexity and cost that this step should mean, it happens that only a sampling of the defects are studied. testing has a cost as well, and earlier screening of weak parts helps reduce the overall wafer cost. A recent project sought to avoid this manual classification of the defects observed at the wafer-inspection level. From a training image dataset, defects are now classified as killers or not-killers; a defect-classification model is built and applied on all the defects; then, for all the parts, a failure probability is assessed per die, from the number of killer or not-killer defects it contains. This means a new capability to screen the dies that are carrying latent defects that could not be screened at electrical step. Machine-learning allows a further step in correlating this failure probability to electrical tests with four types of analysis. The first analysis assesses a failure-probability threshold that is implemented at inspection step and limits the number of parts that should have been electrically tested later: this analysis is at the yield level and saves test cost, beyond quality improvement. The second and third analysis respectively highlights the bins and the tests most correlated with failure probability. The final analysis uses the list of highlighted tests to adjust test limits to screen the parts with failure-probability outliers. These analyses are designed in RStudio and specifically with H2O Driverless AI for the multivariate modeling, and packaged into an RShiny interface. All of them help limit test cost while increasing overall quality.