Prior Knowledge Based Learning Modeling of Emerging Graphene Interconnects

Abstract

With the scaling down of the ICs, need for ‘scaling of interconnects’ is highlighted in terms of considerable RC delay. Traditionally Cu is used for interconnects, however with the sustained miniaturization of integrated circuits (ICs), the effective resistance of Cu increases because of grain boundary scattering and surface boundary scattering. Hence interconnects are becoming the bottleneck in overall performance of the ICs especially in sub 22nm technology nodes. Graphene based materials such as those based on multi-walled carbon nanotubes (MWCNTs) and multilayer graphene nanoribbons (MLGNRs) are being explored as possible replacements of conventional copper interconnects at the on-chip level. In comparison to these two, GNRs being planar in nature offers relative ease of fabrication and have extremely desirable electrical, mechanical and thermal properties for interconnect applications. The MLGNRs are characterized by both the Multi Conductor Circuit (MCC) model and the simplified Equivalent-Single Conductor (ESC) model. Various important processes to enhance the performance of MLGNRs like intercalation doping, insertion of dielectric layer between the GNR layers etc. are studied in detail and efforts were made to incorporate them into modelling. Deterministic Simulations have been carried out for both side contact and top contact MLGNRs for both ESC and MCC model and results have been compared for various signal integrity parameters. Though MLGNR based interconnects can possibly replace the conventional copper interconnects, however, the growth, patterning, and fabrication of such graphene-based interconnects are very sensitive to variations caused due to human / instrumentation errors, difficulty in capturing all the parameters of the system and manufacturing tolerances. Thus, efforts were made to study how the variability in the geometrical, physical and material parameters of MLGNRs affect the signal integrity (SI) performance of these emerging interconnects. Knowledge Based Artificial Neural Network (KBANN) metamodels of dielectric MLGNR were developed and trained using data extracted from SPICE simulations. Various techniques for KBANNs like Prior Knowledge Input (PKI), Source Difference (SD) and Prior Knowledge Input with Source Difference (PKID) were considered during modelling. Also, an artificial neural network (ANN) metamodel using transfer learning approach has been developed for the efficient and much faster statistical signal integrity analysis of MLGNR interconnect networks.