Modeling of Per Unit Length Parameters of Copper-Graphene Hybrid Interconnects using Artificial Neural Networks

Abstract

The necessity of faster responsive chips lead to more wiring density and highly concentrated transistors though the decrement in size of transistors helps in reducing the gate delays of the circuit, the significant reduction in the interconnects lead to colossal increase in the delays surpassing the gate delays too. This effect of increased delays in the copper interconnects is due to scattering effects caused by thinning the copper interconnect lines which also led to other issues such as electromigration reliability. This made us introspect a lot of materials for a viable replacement of copper as interconnect material. Copper-graphene material is one such material which reduces the effects present in the traditional copper interconnects and gives us reasonable delays for the chip interconnect modelling. In order to visualize the interconnect modelling for the copper-graphene interconnects we need Electromagnetic (EM) solver which converts the geometrical into electrical parameters for a sufficiently large samples useful for design space exploration. However, the Electromagnetic (EM) solver though highly accurate is computation intensive and not desirable for large sample model analysis. Machine learning is one of fields which utilizes the data and constructs a map from input to output then used as black box for prediction. Neural networks is one of a technique which must be trained with sufficient data before used as Electromagnetic (EM) solver replacement. Though the time taken for Neural networks is significantly less compared to Electromagnetic solver it still needs to be reduced. Hence, we use Knowledge based Neural networks which uses the data of two models and gives us good prediction with less time. This thesis provides a comprehensive theoretical discussion together with several tutorial application examples, thus complementing the published material.