Abstract
Deep learning (DL) has emerged as a solution for precoding in massivemultiple-input multiple-output (mMIMO) systems due to its capacity to learn thecharacteristics of the propagation environment. However, training such a modelrequires high-quality, local datasets at the deployment site, which are oftendifficult to collect. We propose a transformer-based foundation model for mMIMOprecoding that seeks to minimize the energy consumption of the transmitterwhile dynamically adapting to per-user rate requirements. At equal energyconsumption, zero-shot deployment of the proposed foundation modelsignificantly outperforms zero forcing, and approaches weighted minimum meansquared error performance with 8x less complexity. To address model adaptationin data-scarce settings, we introduce a data augmentation method that findstraining samples similar to the target distribution by computing the cosinesimilarity between the outputs of the pre-trained feature extractor. Our workenables the implementation of DL-based solutions in practice by addressingchallenges of data availability and training complexity. Moreover, the abilityto dynamically configure per-user rate requirements can be leveraged by higherlevel resource allocation and scheduling algorithms for greater control overenergy efficiency, spectral efficiency and fairness.