Physics-informed wireless foundation model
Pretraining and adapting a general-purpose model for communication signals.
Research context
From 2025-06 to 2025-12, I served as an algorithm research intern and project co-lead in the electromagnetic foundation model group at Qiyuan National Laboratory (启元国家实验室), where I worked on a general-purpose model for communication signals.
Pretraining used 80 million signal samples from 14 datasets, with the resulting model adapted to eight communication tasks.
Contributions
- Developed a pretraining and fine-tuning pipeline for communication signals.
- Introduced physics-guided masking as an alternative to random masking.
- Designed physics-inspired modules for a Vision Transformer backbone.
Tasks and results
The model supports tasks including signal denoising, parameter estimation, modulation classification, blind source separation, and radio-frequency identification. Across the evaluated tasks, reported improvements over supervised and self-supervised baselines range from 0.09% to 21.1%.