A Model‐Based Neural Network Pre‐Distorter for Frequency‐Interleaving DAC System

ABSTRACT Frequency interleaving digital‐to‐analog converter (FI‐DAC) can achieve output bandwidths multiple times of single DAC by paralleling multiple DACs. Nevertheless, FI‐DAC contains several errors, such as spectral aliasing, amplitude and phase mismatches, which will cause unwanted spurs or degradation in the output. Pre‐distortion is an effective method for error correction. However, with the increase of paralleled paths of FI‐DAC, the complexity of computing the system frequency response matrix and performing inversion sharply increases, whereas the accuracy sharply decreases, leading to poor performance of existing methods. This paper proposes a neural network pre‐distorter based on gradient descent iterative algorithm for FI‐DAC (PGFN). PGFN iteratively generates the pre‐distortion signal, using learnable parameters to replace the required frequency response matrix and iteration step size, thereby avoiding explicit matrix estimation and accelerating convergence during training. This enables fast and accurate correction of multiple errors in large‐scale FI‐DAC systems. We validate the proposed method on an 8‐path FI‐DAC system. The spurious‐free dynamic range (SFDR) of a multi‐tone signal improves from 12 to 44 dB after pre‐distortion, and the bit error rate (BER) of a PAM‐4 signal decreases from 0.65 to 0.