Physics-Guided Language Model via Low-Rank Adaptation for Path Loss Prediction

The expansion of communication frequency bands and application scenarios necessitates enhanced precision and generalization of wireless channel path loss models. Existing path loss models predominantly rely on measurements within specific frequency ranges and predefined scenarios, and struggle to achieve a balance between accurate prediction and generalization. This paper aims to utilize the reasoning and generalization abilities of Pre-trained Language Models (PLMs) for more effective path loss prediction, particularly in different scenarios and frequencies. Specifically, we introduce theoretical values from the 3GPP propagation models as a prompt to estimate the path loss for different scenarios. The impact of scatterers in the scenario is extracted through advanced vision-language parsing of satellite images to bridge discrepancies between theoretical predictions and empirical measurements, coupled with Low-Rank Adaptation (LoRA) fine-tuning for post-training to achieve cross-modal knowledge transfer. Extensive experiments on real-world datasets show that the proposed model outperforms the state-of-the-art baseline methods. When extrapolated to new scenarios, this scheme exhibits excellent few-shot prediction performance.