A Method based on Gaussian Process Regression for Modelling Burn-in of Semiconductor Devices

Burn-in is a systematic screening method to ensure the reliability of semiconductor devices. It consists in the operation of the manufactured devices under accelerated stress conditions, such as high temperature and voltage. The aim is to remove the devices that would fail in the initial portion of the bathtub curve of the failure rate and estimate the corresponding Early Life Failure Rate (ELFR) value. In practice, performing burn-in is costly and time-consuming, particularly for new technologies. In this context, the present work aims at developing an Artificial Intelligence (AI)-based method to: I) predict the number of defective semiconductor devices within a production lot by resorting to signals measured during the production process; ii) estimate the early life failure probability of the manufactured devices. The method combines: A) dimensionality reduction by Principal Component Analysis (PCA) for the extraction of features characterizing the production process from the measured signals and b) Gaussian Process Regression for the prediction of the number of defective devices in the lot. The method is applied to artificial data simulated to emulate burn-in process data. The obtained results show a satisfactory accuracy in the prediction of the number of defective devices in the lot and of the corresponding early life failure probability.