End-to-End Learning for Time-Varying Non-Binary Molecular Communications

This paper presents a novel end-to-end learning approach for non-binary transmission in molecular communication (MC) systems, addressing both inter-symbol interference (ISI) and concentration-dependent noise effects. The proposed autoen-coder (AEC) generates non-uniform multi-level concentration shift keying (CSK) modulation at the transmitter (Tx), which can be detected with minimal error at the receiver (Rx). To mitigate the effects of concentration-dependent noise more effectively, the loss function is augmented with a penalty term based on the maximum concentration level generated at the Tx, resulting in improved power efficiency and lower error probability. Furthermore, the paper introduces conditional feature-wise linear modulation (FiLM) layers in the AEC to enhance adaptability in time-varying MC scenarios, where channel parameters such as the diffusion coefficient vary over time.