ChannelGPT: A Large Model toward Real-World Channel Foundation Model for 6G Environment Intelligence Communication

Environment intelligence communication (EIC) is envisaged to integrate and utilize sensing and artificial intelligence (AI) abilities brought by 6G, to acquire dynamic real-world environment information and predict wireless channel fading online for the environment-adaptive system optimization. As the cornerstone of EIC, channel fading model is a vital procedure for 6G high-dynamic and diverse scenarios, whose accuracy, adaptability and generalization intrinsically determine the ultimate system performance. However, traditional specialized AI models are tailored to a particular task or a specific scenario, leading to great challenges especially for the diverse, complicated and high-dynamic wireless channel. In this article, we first propose ChannelGPT, a pretrained large model, which is designed for real-world channel generation and prediction by fine-tuning only a few parameters with multimodal data of the new environment. To capture the real-world environment features, multimodal data such as image, point cloud and position information from diverse environments are utilized to pretrain ChannelGPT with specifically tailored preprocessing and embedding layer. Then, zero-shot or few-shot learning is employed, and ChannelGPT demonstrates rapid adaption to multitask channel prediction including long-range prediction, multimodal sensing and real-world generalization. In essence, these properties enable ChannelGPT to build general-purpose capabilities to generate multi-scenario channel parameters, associated environment maps and wireless knowledge simultaneously, in terms of each task requirement. Moreover, preliminary results of ChannelGPT demonstrate the accuracy, efficiency and generalization ability towards real-world channel foundation model. Finally, open issues and challenges are outlined.