Neural network-based modeling and design of on-chip spiral inductors

Abstract

A neural network approach is presented for the modeling and re-design of high-Q on-chip spiral inductors. The approach involves the creation of neural network models to map 3D multi-level spiral inductor geometric and material characteristics to SPICE equivalent circuit parameters. The neural network replaces computationally expensive FEM-based extraction and field solution. The approach is especially attractive because it is capable of accurately and efficiently predicting important inductor characteristics such as self-inductance, Q-factor, self-resonant frequency and parasitic resistance and capacitance. It also offers substantial computational savings over field solution-evaluation of neural model required on average 2% of the cpu time required for field solution. The neural approach served not only as a basis for fast spiral inductor circuit extraction but also permits fast spiral layout design refinement from post-optimization inductor circuit-level parameters.