Using Polynomial Regression and Artificial Neural Networks for Reusable Analog IC Sizing

In this paper, the use of machine learning techniques to repurpose already available Pareto optimal fronts of analog integrated circuit blocks for new contexts (loads, supply voltage, etc.) is explored. Data from previously sized circuits is used to train models that predict both circuit performance under the new context and the corresponding device sizes. A two-model chain is proposed, where, in the first layer, a multivariate polynomial regression estimates the performance tradeoffs. The output of this performance model is then used as input of an artificial neural network that predicts the device sizing that corresponds to that performance. Moreover, the models are trained with optimized sizing solutions, leading almost instantly to predicted solutions that are near optimal for the new context. The proposed methodology was integrated into a new framework and tested against a real circuit topology, with promising results. The model was able to predict wider and, in some cases, better, performance tradeoff, when compared to independent optimization runs for the same context, despite requiring 400 times fewer circuit simulations.