Neural Transmitter with Co-Optimized Digital Front End Functions

Abstract

In order to increase Power Amplifier (PA) linearity and efficiency in the transmission chain wireless communication systems, Digital Front-End (DFE) functions such as Crest Factor Reduction (CFR), Hard Limiter (HL) and Digital Pre-Distortion (DPD) are typically built and adjusted independently. Despite the effectiveness of this modular strategy, it does result in a high hardware footprint. The use of Neural Networks (NNs) to simultaneously model and co-optimize DFE functions for wireless communication systems with memory-effect-exhibiting PAs is examined in this thesis.

As a reference, a functional hardware-compatible legacy DFE with CFR, HL and DPD was developed, which was used to evaluate the NN-based DFE solutions. The NN models were compressed using pruning, projection, and quantization methods to allow deployment in contexts with limited resources. To compare the hardware cost of these solutions, bit-exact models were developed. The results demonstrate that, although at a much greater hardware cost, uncompressed NN-based DFE models give better performance than the legacy DFE. Implementation complexity is significantly reduced by compression, though the compression techniques investigated also gives significant performance loss. The quantization-based compression give the best complexity reduction, though the projection-based compression offers a better performance-complexity trade-off than quantization.

Everything being considered, the findings show the promise of NN-based co-optimization of DFE functions while emphasizing the necessity of better compression methods for realistic implementation.

Popular Abstract

Radio signals carrying information are sent and received all around us in wireless communication systems, between cellphones and cell towers, and vice versa. In order for a signal to reach its intended receiver it must be sent with sufficient power. For this reason the signal power is amplified using a Power Amplifier (PA). This amplification consumes a lot of power and is nonlinear, meaning that different parts of the signal are amplified at different strengths, making it difficult for the receiver to decipher the original signal information.

To help the PA run energy efficiently and amplify the signal power as equally as possible across the signal, a so called the Digital Front-End (DFE) is applied to modify the signal before amplification. Traditionally the DFE consists of several components each handling different aspects of the signal modification. This solution works well for 4G and 5G networks, but becomes increasingly more expensive requiring more resources to comply with the tougher signal requirements of the 6G networks currently under development.

As such we investigated replacing the components of the DFE with Artificial Intelligence (AI), specifically with an AI model called Neural Networks (NNs) inspired by the human brain. NNs are known to handle complex problems well, given the right training, in which the NN starts out knowing nothing and is given a huge amount of examples to practice on and gets better by learning from its mistakes. NNs are also known to consume a lot of resources, both during training and when using them. So we investigated how to train an NN with the best performance while consuming the least resources. Our hope was that replacing the DFE components with an NN would either help modify the signals in a better way making PA work better, or in a more efficient way reducing resource usage.

Our investigations led us to NN-based DFE solutions which either modified the signals in a better way but with much higher resource usage than for the traditional solutions, or used about the same amount of resources as the traditional solutions, but failed to properly modify the signals. There is still a lot to explore regarding the design and training of these NN-based DFE solutions, and we see a lot of potential for future investigations.

• author: Hofvander, Astrid and Attupurath, Noopura Parvathi
• supervisor: Monika Eisenmann
• organization: Mathematics (Faculty of Engineering)
• alternative title: Neural Transmitter
• course: FMAM05 20261
• year: 2026
• type: H2 - Master's Degree (Two Years)
• subject: Technology and Engineering; Mathematics and Statistics
• keywords: Neural Network (NN), Wireless Communications, Power Amplifier (PA), Digital Front-End (DFE), Digital Predistortion (DPD), Crest Factor Reduction (CFR), Hard Limiter (HL)
• publication/series: Master’s Theses in Mathematical Sciences
• report number: LUTFMA-3636-2026
• ISSN: 1404-6342
• other publication id: 2026:E112
• language: English
• additional info: The thesis project was conducted jointly with Astrid Hofvander from Lund University and Noopura Parvathi Attupurath from Chalmers University of Technology as part of a collaborative project with Ericsson, and is also being published by Chalmers University of Technology
• id: 9250244
• date added to LUP: 2026-09-17 08:26:26
• date last changed: 2026-09-17 08:26:26