D3.5 Report on End-to-End Air-Interface Learning

This report explores the feasibility and practical relevance of end-to-end learning approaches for next generation 6G air interfaces. Our research focuses on three key innovations: pilotless communication systems, scalable symbol modulation learning, and joint source channel coding for short packet transmissions. The pilotless communication system eliminates traditional reference signals by jointly training a neural receiver with custom trainable constellations, embedding channel estimation mechanisms implicitly within the transmitted data. This approach demonstrates competitive block error rates while achieving up to 8% higher goodput compared to 5G NR baseline implementations due to a reduced piloting overhead. We further introduce a scalable autoencoder structure capable of supporting any M-ary modulation through a single AI/ML model, with robustness against non-linear impairments as expected in future high-frequency communications. For short packet transmissions, we propose joint source channel coding and modulation (JSCCM) mechanisms specifically optimized for compressed CSI feedback, addressing implementation challenges like an increased peak-to-average power ratio (PAPR) when multiplexing with other logical channels. These innovations require minimal modifications to existing infrastructure, providing a practical pathway toward intelligent, reconfigurable physical layers for future wireless systems while maintaining backward compatibility with current standards.