Learned Beamforming

Treats the channel matrix as a 2D image, applying convolutional filters to extract local spatial features. This architecture excels at learning structured spatial correlations in uniform planar arrays.

• Input: Channel state information (CSI) matrix reshaped into a 2D grid
• Mechanism: 2D convolutions capture local antenna correlations; fully connected layers output precoding vectors
• Key Advantage: Parameter efficiency through weight sharing; naturally handles grid-structured arrays
• Use Case: Hybrid precoding for mmWave massive MIMO with uniform rectangular arrays

Models the wireless network as a graph where antennas or users are nodes and interference channels are edges. GNNs learn optimal beamforming vectors through iterative message passing between nodes.

• Input: Graph representation of the MIMO interference channel
• Mechanism: Node embeddings are updated by aggregating neighbor information; permutation equivariance ensures scalability to varying network topologies
• Key Advantage: Generalizes to unseen network sizes without retraining; inherently captures interference topology
• Use Case: Distributed beamforming in multi-cell networks with dynamic user counts

Unfolds a classical iterative optimization algorithm, such as WMMSE or ADMM, into a fixed number of neural network layers. Each layer corresponds to one iteration, with learnable parameters replacing hand-tuned step sizes.

• Input: CSI matrix and initial precoder guess
• Mechanism: A truncated iterative algorithm is mapped to a feedforward network; backpropagation optimizes the algorithm's hyperparameters end-to-end
• Key Advantage: Combines the interpretability and structure of model-based optimization with the speed of learned inference
• Use Case: Real-time WMMSE-based precoding where 3-5 unfolded layers match 100+ classical iterations

Applies the self-attention mechanism to the sequence of channel vectors across subcarriers or antennas. The transformer learns long-range dependencies in frequency-selective channels for wideband beamforming.

• Input: Sequence of CSI vectors across OFDM subcarriers
• Mechanism: Multi-head self-attention computes pairwise interactions between all subcarriers; the output is a beamforming matrix for the entire band
• Key Advantage: Captures frequency-domain correlations that CNNs miss; superior performance on wideband, frequency-selective channels
• Use Case: Subcarrier-level precoding in 5G NR OFDM systems with high delay spread

Processes a temporal sequence of CSI estimates to predict beamforming vectors for mobile users. The recurrent structure maintains a hidden state that tracks channel evolution, eliminating the need for explicit channel prediction.

• Input: Time series of past CSI estimates or received pilots
• Mechanism: LSTM or GRU cells maintain a latent channel state; the output layer maps the hidden state to beamforming weights for the next time slot
• Key Advantage: Implicitly learns channel aging dynamics; robust to Doppler spread without a separate Kalman filter
• Use Case: Beam tracking for vehicular communication in high-mobility V2X scenarios

A multi-user MIMO transceiver implemented as a single end-to-end autoencoder. The transmitter network learns a joint precoding and modulation mapping, while the receiver network learns joint detection, all optimized over a stochastic channel model.

• Input: Bit streams for multiple users at the transmitter; received signals at the receiver
• Mechanism: Both transmitter and receiver are neural networks trained jointly via backpropagation through a differentiable channel model
• Key Advantage: Discovers non-intuitive precoding structures that outperform classical linear methods like zero-forcing on non-Gaussian channels
• Use Case: Joint precoding and detection for non-linear channels with hardware impairments

An end-to-end neural network that jointly optimizes spatial multiplexing and beamforming. Unlike a standalone learned beamformer, a MIMO autoencoder learns the mapping from bit streams directly to antenna elements, implicitly discovering optimal precoding strategies without explicit channel state information (CSI) feedback. This collapses the traditional separation between coding, modulation, and beamforming into a single, differentiable optimization problem.

A neural compression framework critical for Frequency-Division Duplex (FDD) massive MIMO. The user equipment (UE) uses an encoder network to compress the downlink channel matrix into a low-dimensional codeword. The base station (gNB) uses a decoder network to reconstruct the CSI. This learned approach dramatically reduces uplink feedback overhead compared to traditional codebook-based methods like Type-II CSI reporting, directly enabling high-resolution learned beamforming at the transmitter.

Neural beamforming is only as good as the channel state information it receives. Channel Estimation AI replaces classical estimators like Least Squares (LS) and Minimum Mean Square Error (MMSE) with deep neural networks that can learn complex channel statistics and denoise pilot signals. In high-mobility scenarios, these networks perform predictive beamforming by forecasting future channel states, allowing the beamformer to proactively steer energy before the channel changes.

A paradigm where the mathematical aggregation of local model updates from distributed devices is computed directly via the superposition property of the wireless multiple-access channel. Learned beamforming is essential here: the receiver must design a combining vector that aligns the analog gradient transmissions from multiple devices to compute a weighted sum. This turns interference into a computation resource, drastically reducing communication latency for distributed edge learning.

In millimeter-wave (mmWave) systems, fully digital beamforming is power-prohibitive. Hybrid beamforming splits precoding between a low-dimensional digital baseband and a high-dimensional analog phase-shifter network. Learned beamforming models are increasingly used to jointly optimize this split, using neural networks to map channel estimates to both analog beamsteering codebook indices and digital baseband precoder matrices, solving a non-convex optimization problem in a single inference pass.

An end-to-end learned transmitter-receiver pair optimized to maximize the mutual information with a legitimate receiver while minimizing information leakage to an eavesdropper. The learned beamformer acts as a spatial filter, steering a high-SNR beam toward the intended user while creating a null or artificial noise toward the eavesdropper's direction. This learns a joint encryption and precoding scheme without a pre-shared key, using the channel's uniqueness as a security primitive.