Additive Neural Network Based Static and Dynamic Distortion Modeling for Prior-Knowledge-Free Nyquist ADC Characterization

This paper presents a prior-knowledge free modeling method for Nyquist ADCs. Current ADC modeling methods mainly base on known circuit implementation and non-idealities, thus hard to recover non-linear static and dynamic distortions. The proposed method adopts an additive neural network with binary inputs to achieve a data driven, prior-knowledge free modeling method. Both static and dynamic distortions are modeled by two separate sub-network. Also, a batch generation scheme is used to enhance the noise insensitivity, facilitating small sample training, when only simulation results are available. The proposed methods are validated by three typical non-ideal ADC designs, including a SAR ADC with capacitor mismatch, an ultra-high speed ADC with NMOS sampling switch, and a SAR ADC with a bandwidth limited reference source. All the non-linearity and FFT spectrum plots show the proposing model can accurately model both static and dynamic distortion with less than 1dB spur mismatch.