Fast-Accurate Non-Polynomial Metamodeling for Nano-CMOS PLL Design Optimization

At the nanoscale domain, the simulation, design, and optimization time of the circuits have increased significantly due to high-integration density, increasing technology constraints, and complex device models. This necessitates fast design space exploration techniques to meet the shorter time to market driven by consumer electronics. This paper presents non-polynomial metamodels (surrogate models) using neural networks to reduce the design optimization time of complex nano-CMOS circuit with no sacrifice on accuracy. The physical design aware neural networks are trained and used as metamodels to predict frequency, locking time, and power of a PLL circuit. Different architectures for neural networks are compared with traditional polynomial functions that have been generated for the same circuit characteristics. Thorough experimental results show that only 100 sample points are sufficient for neural networks to predict the output of circuits with 21 design parameters within 3% accuracy, which improves the accuracy by 56% over polynomial metamodels. The generated metamodels are used to perform optimization of the PLL using a bee colony algorithm. It is observed that the non-polynomial (using neural networks) metamodels achieve more accurate results than polynomial metamodels in shorter optimization time.