End-to-End Waveform Design for Nonlinear Satellite Links with a Convolutional Autoencoder

We present a convolutional autoencoder (CAE) for single-carrier (SC) waveform design in satellite links with nonlinear high-power amplifiers (HPAs). The CAE jointly optimizes bit-error rate (BER), peak-to-average power ratio (PAPR), and adjacent-channel power ratio (ACPR) via an augmented-Lagrangian objective. Using measured block-upconverter (BUC) data with AM/AM and AM/PM distortions across temperature and frequency variations, we show consistent improvements over conventional PAPR-reduction techniques and baselines, with strong generalization across frequency and temperature variations. Experiments with 64-QAM demonstrate BER and spectral advantages at practical back-off levels without explicit digital predistortion (DPD).