Technology in action

Wireless AI Research

Research: AI in mobile communications

A neural receiver promises even more stable data transmission for mobile radio. Together, NVIDIA and Rohde & Schwarz are developing a test setup.

27-Mar-2025

Wireless transmissions are always subject to interference. To compensate, today’s mobile communications rely on channel estimation and equalization. The transmitter sends additional pilot signals known on the receiving end along with the actual data stream. If any distortions are detected in the pilots when they arrive at the receiving end, a digital signal filter equalizes them. Powerful signal processing algorithms calculate the right filter parameters based on the degree of distortion in the pilot signals.

These interference suppression methods account for different conditions that commonly occur in mobile communications. Interference in a wireless transmission sent by someone riding a bike in the countryside will be different than one from a crowded pedestrian area or on a moving train.

Where artificial intelligence comes in

Every successful data connection to a mobile device is proof positive of how sophisticated current signal processing already is, but there are still limits. Optimization methods are never perfect because signal processing algorithms are developed on the basis of standardized channel profiles – assumed models that only give an approximation of actual operating conditions. Training AI models with data sets that better reflect real-world conditions would enable more effective methods for channel equalization, and thus more stable wireless connections with higher data throughput.

A permanent fixture in mobile communications

Figure 1 shows the specific approach NVIDIA is taking with neural receivers: In the receiver (RX), the signal processing block for channel estimation, channel equalization and demapping is replaced by a trained machine model that handles all three tasks. The neural receiver was developed using NVIDIA's Sionna open source software library, which is specifically designed for 5G and 6G research.

Comparison of classic transmitter setup and AI-based transmitters.

In a neural receiver (below), a trained machine learning model handles channel estimation, equalization and demapping. In the classic linear minimum mean squared error (LMMSE) receiver architecture shown above, deterministic software algorithms perform these tasks.

/

The computing power needed to effectively train a neural receiver is still very high and requires graphical processing units (GPUs). However, initial findings indicate that the results justify the expense. Researchers are also optimistic that the AI models can be trained with quite a bit less computing power. Many experts now assume that AI models will become a permanent fixture in 6G mobile communications signal processing.

Whether there is enough high-quality, real world training data is still an open question. The need for actual data from the field will increase as the market matures. Up to now, synthetic training data from simulations or generated data sets have been perfectly adequate in the current phase of research.

Measurement technologies to evaluate performance

Rohde & Schwarz already provides the right signal sources and signal analysis tools needed to set up a test environment for the neural receiver. The R&S®SMW200A vector signal generator emulates individual users transmitting signals in MIMO signal configuration and adds noise and fading as needed to simulate realistic radio channel conditions. The receiver in the current test setup is the R&S®MSR4 universal satellite receiver with four parallel receiving channels. It forwards the signals through a real-time streaming interface to a server, where the R&S®Vector Signal Explorer (VSE) software synchronizes the signals and performs a fast Fourier transform (FFT). This FFT dataset is then used as the input for the neural receiver.

To assess quality, the reconstructed data blocks are compared with the original data. Calculating the ratio of data blocks with errors to the total number of transmitted data blocks yields the block error rate (BLER).