Deep Learning-based DMRS Configuration for MIMO Channel Estimation

This paper studies the application of tools from Artificial Intelligence and Machine Learning (AI/ML) for the adaptive configuration of reference signals (pilots). Specifically, we propose a deep learningbased framework to infer on the configuration of userspecific demodulation reference signals (DMRS) that are used for composite channel estimation (CCE) in multiple-input multiple-output (MIMO) systems. A proof-of-concept implementation is provided here, where a neural network engine residing at the terminal is trained offline through synthetic 5G New Radio (NR) standards-compliant waveforms and channel models. The evaluation results highlight the gains that can be achieved through adaptive DMRS configuration, particularly in terms of reduced reference signal overhead.