Reliability characterization and prediction of high k dielectric thin film

As technologies continue advancing, semiconductor devices with dimensions in nanometers have entered all spheres of human life. This research deals with both the statistical aspect of reliability and some electrical aspect of reliability characterization. As an example of nano devices, TaO_x-based high k dielectric thin films are studied on the failure mode identification, accelerated life testing, lifetime projection, and failure rate estimation. Experiment and analysis on dielectric relaxation and transient current show that the relaxation current of high k dielectrics is distinctive to the trapping/detrapping current of SiO₂; high k films have a lower leakage current but a higher relaxation current than SiO₂. Based on the connection between polarization-relaxation and film integrity demonstrated in ramped voltage stress tests, a new method of breakdown detection is proposed. It monitors relaxation during the test, and uses the disappearing of relaxation current as the signal of a breakdown event. This research develops a Bayesian approach which is suitable to reliability estimation and prediction of current and future generations of nano devices. It combines the Weibull lifetime distribution with the empirical acceleration relationship, and put the model parameters into a hierarchical Bayesian structure. The value of the Bayesian approach lies in that it can fully utilize available information in modeling uncertainty and provide cogent prediction with limited resources in a reasonable period of time. Markov chain Monte Carlo simulation is used for posterior inference of the reliability projection and for sensitivity analysis over a variety of vague priors. Time-to-breakdown data collected in the accelerated life tests also are modeled with a bathtub failure rate curve. The decreasing failure rate is estimated with a non-parametric Bayesian approach, and the constant failure rate is estimated with a regular parametric Bayesian approach. This method can provide a fast and reliable estimation of failure rate for burn-in optimization when only a small sample of data is available.