Calibration of Gain Mismatches in TIADCs using Convolutional Neural Networks

Time-interleaved analog-to-digital converters (TIADCs) are widely used in high-speed sampling systems, but their performance is degraded by channel mismatches among sub-ADCs. This paper presents a lightweight one-dimensional convolutional neural network (1-D CNN)-based calibration method for compensating gain mismatches in TIADCs. The proposed CNN is trained in a supervised manner to estimate the gain-mismatch-induced error component from the distorted TIADC output. During training, the target error is generated from the difference between the gain-mismatched TIADC output and a reference mismatch-free ADC output. During inference, only the TIADC output is required, and the calibrated signal is reconstructed by subtracting the estimated error from the mismatched output. Simulation results for 4-channel and 8-channel TIADC systems show that the proposed method effectively suppresses gain-mismatch-induced spurious tones. For the 4-channel TIADC, the proposed calibration improves SNDR by 31.23 dB and SFDR by 61.44 dB. For the 8-channel TIADC, improvements of 32.43 dB in SNDR and 61.02 dB in SFDR are achieved. The proposed CNN contains approximately 17k trainable parameters, suggesting its potential for low-complexity digital background calibration.