Environment-Aware Beamforming Prediction: A CKM-Inspired Approach

Abstract

In multiple-input multiple-output (MIMO) systems, beamforming is a core enabler for concentrating signal power toward targeted users to enhance spectral efficiency and network capacity. However, realizing the full potential of beamforming critically relies on accurate channel state information (CSI). Conventional CSI acquisition requires extensive pilot transmissions and continuous feedback, incurring high overhead. To overcome this limitation, inspired by the novel concept of channel knowledge map (CKM), we propose an environment-aware beamforming prediction (EA-BFP) framework that directly maps environmental topology and transceiver locations to optimal beamforming vectors. The framework employs a dual-branch neural network: a modified ResNet extracts spatial features from the environmental map, while fourier position encoding (FoPE) provides precise high-dimensional coordinate embedding, which enables zero-shot beam decisions. Simulation results demonstrate that our approach achieves lower prediction errors and higher achievable rates than baseline models, showcasing superior main lobe alignment, data efficiency, and cross-scenario generalization under limited fine-tuning samples.