Kriging Bootstrapped Neural Network Training for Fast and Accurate Process Variation Analysis

Oghenekarho Okobiah,  Saraju Mohanty,  Elias Kougianos
University of North Texas, Denton


Speeding up the design optimization process of Analog/Mixed-Signal circuits has been a subject of active research. Techniques such as metamodeling, artificial neural networks, and optimization over SPICE netlists have been used. While the results are accurate and promising, the effects of process variation on design space exploration still persist. Metamodels created by existing techniques are still not variation aware. This paper presents a novel technique for fast and accurate process variation analysis of nanoscale circuits. The technique combines traditional Kriging with an artificial neural network (ANN) to achieve the objective. Kriging captures correlated process variations of the circuits and accurately trains the ANN to generate the metamodels. The proposed technique uses Kriging to bootstrap target samples used for the ANN training. This introduces Kriging characteristics, which account for correlation effects between design parameters, to the ANN. As a case study of the proposed method, Kriging bootstrapped trained ANN metamodels are presented for an 180 nm Phase-Locked Loop (PLL).