Transformer Based Channel Identification

Once an electrical system completes assembly, highspeed channel characterization is complex as access points are limited, and probing by a network analyzer is not a possibility. Further active channels may now have different types of I/O from NRZ to PAM4 levels, adding additional complications. This work describes a method for predicting a channel's behavior and implicitly learning the underlying system parameters in two stages. The first stage uses an autoregressive transformer structure to predict the channel's behavior as a single-bit response (SBR). The second stage uses the model's latent space and unsupervised dimensionality reduction to determine the underlying system parameters. The results show that the model can predict the SBR for completely unseen channels for both NRZ and PAM-4 despite the model training using only NRZ waveforms. Further, unsupervised dimensionality reduction enables the determination of system parameters, specifically the length of the overall transmission line.