End –to –End Learning for Hardware Impaired Downlink Cell –Free Massive MIMO

Cell-free massive multiple-input multiple-output (CF-mMIMO) promises macrodiversity and high spectral efficiency through cooperation among many distributed access points (APs). However, hardware imperfections such as power amplifier (PA) nonlinearities become a limiting factor when many low-cost radio-frequency chains operate in parallel. Conventional mitigation techniques either rely on accurate distortion models or require expensive digital pre-distortion (DPD) at each transmitter. In this paper we propose a hardware-aware end-to-end (E2E) learning framework for the downlink of a CF-mMIMO network. The transmitter employs a neural network (NN) encoder hosted at the central processing unit that encodes each user data symbol. These symbols are then locally precoded by a full-pilot zero-forcing (FZF) matrix at each AP and transmitted before being received and classified by a second NN decoder at each user terminal. The encoders and decoders are trained jointly in an offline phase to minimize the categorical cross-entropy loss on a large dataset of randomly drawn quadrature amplitude modulation (QAM) symbols and channel realizations. Once trained, the solution operates without requiring any knowledge of the PA model. Simulation results show that the proposed E2E framework achieves lower symbol error rate (SER) than a conventional FZF precoder with and without DPD, especially in the hardware-impairment-dominated regime. Furthermore, the proposed architecture is scalable and operates without backpropagation over the fronthaul during inference.