AI/ML Empowered High-Order Modulations for 6G High Capacity Communications

We present new high-order modulation schemes for reliable and high capacity 6G commutations, which beat the state-of-the-art ones adopted in the latest cellular and broadcast communication standards. The performance of modulation schemes is mainly determined by the modulation symbol constellation and the mapping of the encoded bits to a modulation symbol. The square QAM schemes of 16/64/256/1024 modulation orders have been widely adopted in various communication standards, however they fundamentally exhibit up to 1.53dB loss, called shaping loss, in terms of the required SNR for a target date rate with respect to the Shannon capacity bound. In this work, we develop machine-learning (ML) based modulation optimization methodologies and present optimal modulation schemes (symbol constellation, bit-to-symbol mapping) for various modulation orders and SNRs. The neural network architecture and the training methods are designed taking into account the desired properties of well performing modulations. This significantly helps the training converge to a seemingly global optimal state and result in the modulation constellations and bit-to-symbol mappings which reduce the shaping loss to the Shannon capacity bound to a large extent. The new high order modulation schemes obtained through the ML methodologies outperform the state-of-the-art modulations adopted in the ATSC (Advanced Television Systems Committee) 3.0 standards, at least a few dB in case of 1024-ary modulations with LDPC coding, and will enable achieving higher reliability and capacity for 6G communications.