93 Efficient and Accurate Neural Network-Based Macro-Models for Spiral Inductors

A neural network approach is presented for efficient and accurate model parameter extraction for integrated circuit spiral inductors. The approach involves the creation of neural network models to map spiral inductor geometric characteristics to s- and y- parameters over the frequency range 0.1GHz–100GHz. The objective is to develop a more efficient replacement model for spiral inductors — a circuit element for which a detailed simulation model already exists — i.e. a replacement neural macro-model. This neural model is achievable without sacrificing much accuracy — in fact, given that designers try to create near ideal circuit behavior (e.g. approximately linear) using devices that are themselves far from idea! (highly non-linear), a “simplified” replacement macro-model for circuit simulation is a very reasonable proposition for gaining simulation efficiency. The approach is especially attractive because it is capable of modeling skin and proximity effects as well as changes in effective series resistance at higher frequencies. Additionally, the neural network macro-model can predict important inductor characteristics such as Q-factor and self-resonant frequency. Substantial computational savings over detailed circuit simulation are available — ranging from 88 % to 95%. These ensure that the designer can rapidly evaluate the performances of alternative spiral geometries over a wide frequency range without the need for prototyping. What are the advantages of using AI methods to develop neural-network based macro-models? There are several potential advantages. One is improved anticipation of the consequences of technological choices allowing early shifts in processing methods and device∕circuit design well before actual physical manufacture. Neural network models can result in reduced time-to-market especially for RFICs in promising new technologies. Another advantage of the neural network model is its mathematical property of differentiability — important for efficiency of higher order time-integration schemes particularly in analog circuit simulators. Neural network models can be made infinitely smooth, hence differentiable resulting in more realistic results for certain types of analyses. Finally, a third advantage of the neural network is the ability to cany out “reverse mapping” by permitting calculation of optimal spiral geometry for a given set of desirable s-parameters and Q factor at specified frequencies.