Unlocking the Potential of Deep Learning-Based Approaches in Delta Sigma Modulators

This paper presents a new approach to address the design challenges of all-digital transmitters (ADTs) in software-defined radios (SDRs) by introducing a deep learning (DL)-based delta-sigma (Δ∑) modulator. Different state-of-the-art DL structures are adopted to appropriately select the best fit for this processing function, including deep neural networks (DNN), convolutional neural networks (DNN), recurrent neural networks (RNN), and long short-term memory (LSTM). This alternative design approach eliminates the dependency between the sampling frequency and the number of parallel cores, as well as between the noise shaping and the Δ∑ core design. The proposed DL-based Δ∑ algorithm is evaluated across different oversampling ratios and modulation orders, allowing for the extraction of key insights and verification steps to support future research on DL-based Δ∑ modulators.