Wireless AI Evolution: From Statistical Learners to Electromagnetic-Guided Foundation Models

Abstract

While initial applications of artificial intelligence (AI) in wireless communications over the past decade have demonstrated considerable potential using specialized models for targeted communication tasks, the revolutionary demands of sixth-generation (6G) networks for holographic communications, ubiquitous sensing, and native intelligence are propelling a necessary evolution towards AI-native wireless networks. The arrival of large AI models paves the way for the next phase of Wireless AI, driven by wireless foundation models (WFMs). In particular, pre-training on universal electromagnetic (EM) principles equips WFMs with the essential adaptability for a multitude of demanding 6G applications. However, existing large AI models face critical limitations, including pre-training strategies disconnected from EM-compliant constraints leading to physically inconsistent predictions, a lack of embedded understanding of wave propagation physics, and the inaccessibility of massive labeled datasets for comprehensive EM-aware training. To address these challenges, this article presents an electromagnetic information theory-guided self-supervised pre-training (EIT-SPT) framework designed to systematically inject EM physics into WFMs. The EIT-SPT framework aims to infuse WFMs with intrinsic EM knowledge, thereby enhancing their physical consistency, generalization capabilities across varied EM landscapes, and overall data efficiency. Building upon the proposed EIT-SPT framework, this article first elaborates on diverse potential applications in 6G scenarios of WFMs, then validates the efficacy of the proposed framework through illustrative case studies, and finally summarizes critical open research challenges and future directions for WFMs.

Figures (5)

• Figure 1: Development of wireless AI and electromagnetic information theory.
• Figure 2: Electromagnetic information theory: (a) Field-based, (b) Circuit-based, and (c) Wavenumber-domain methods.
• Figure 3: Self-supervised pre-training methods for WFMs. Generative pre-training aims to reconstruct the original input from a latent representation. Conversely, contrastive methods learn a feature space organized by similarity, by pulling similar samples closer while pushing dissimilar ones apart. The hybrid generative-contrastive pre-training framework simultaneously integrates the reconstruction and contrastive tasks.
• Figure 4: The proposed EIT-SPT framework for WFMs. It systematically injects physical laws into the WFM lifecycle through three synergistic layers: 1) The EM-compliant data genesis layer ensures the input follows physical laws; 2) The physics-informed architecture layer provides the necessary structural inductive bias for continuous fields; and 3) The self-supervised pre-training layer employs physics-rooted tasks to enforce causal learning.
• Figure 5: Case studies for channel estimation and positioning of the proposed EIT-SPT framework. For the channel estimation case, we consider a multi-user HMIMO system, where a base station comprises densely packed uniform planar array antennas with sub-wavelength spacing. For the wireless positioning case, the CSI dataset is measured via a real-world distributed Massive MIMO testbed, where the total grid of positioning area spans 1.25 m by 1.25 m 9129126.