Machine Learning based Prior-Knowledge-Free Nyquist ADC Characterization and Calibration

Abstract

Traditional ADC characterization and calibration algorithms, for example, least mean square, was based on sophisticated ADC modeling knowledge. The recent artificial intelligence surge motivates us to develop a general and prior-knowledge-free method suitable to a variety of ADC distortion sources. This paper describes such a method using neural network. The input of the network is bit-wisely encoded, and the network utilizes a residual topology for fast convergence. A number of Nyquist ADC prototypes, including an open-loop amplifier based pipelined SAR ADCs, an NMOS based sampling circuit etc., are characterized or calibrated by adopting the machine learning based methods. Measurement results shows the improvement is nearly 30dB SFDR.