Research On The Interconnect Reliability Using An Automated Model Generation Algorithm

With the continuous size scaling of CMOS transistors, current densities in interconnects increase all the way, which results in the migration of metal atoms in interconnects. The difference in the number of metal atoms moving in and out of a unit volume per unit time results in an atomic flux divergence(AFD) in the metal line, and hence the formation of the voids and hillocks, which leads to the interconnect failure. The electromigration(EM) failure occurs due to the gradual displacement of the metal atoms in interconnects. Therefore, EM has become a serious issue in ultra large scale integration(ULSI) circuits during the IC design for reliability.To evaluate EM of interconnects, finite element models(FEMs) are used to calculate AFD. The main driving forces for AFD are the electron wind force(EWF), the stress gradient induced driving force(SGIDF), and the temperature gradient induced driving force(TGIDF). However, the FEM simulation is very time-consuming, especially for complex ICs. When the layout or the interconnect structure changes, the 3-D finite-element model must be reconstructed and simulated manually, which is very tedious and time consuming. Therefore, the existing analytical methods can’t predict the performance of FEMs in a shorter time.To speed up the FEM simulation in the analysis of interconnect reliability, Artificial Neural Networks(ANNs) are chosen to obtain the solution of EM because it can be trained to learn the high nonlinearity behaviors of ICs.Based on ANN, an advanced algorithm for automated model generations(AMGs) is presented to calculate the AFD of metal interconnects. The proposed algorithm automates data generations, determinations of data distributions, model structure adaptations, and the model training in a systematic framework. It can further reduce the amount of training data via an adaptive sampling process, and shorten the period of model developments. To verify the high efficiency of the proposed approach, a CMOS inverter, as an example, is built in a finite-element environment. A comprehensive comparison between AMG and the conventional manual modeling approach in the reliability analysis of interconnects is presented. The result shows that there is a running time saving up to 51.3% by using AMG while maintaining high accuracy without a waste of physical labor. Besides, the behavior of the interconnect reliability with respect to temperature, current, interconnect materials and sizes are analyzed. Sensitivity analysis is performed to find out the parameter to which AFD is the most sensitive. Monte Carlo analysis is also performed to predict the relationship between process variations and interconnect reliability in manufacturing, which can be a guidance for IC design for reliability.Above all, this paper focuses on the saving of time in the analysis of interconnect reliability by using AMG. Besides, AMG can automatically build models by invoking ANSYS so that the complex operations in ANSYS are avoided.