Generative AI-Empowered Semantic Twin Channel Model for ISAC
Yi Chen, Yatao Hu, Ming Li, Chong Han
TL;DR
The paper addresses the ISAC channel modeling gap where sensing-relevant multipath is often lost in conventional models. It introduces environmental semantics as a unifying abstraction and presents a generative AI–powered semantic twin channel model (STCM) that yields ensembles of semantically conditioned, physically plausible channel realizations via a four-stage pipeline, including a physics-grounded synthesizer $M(\boldsymbol{\theta})$ and a conditional generator $G(s,z)$. It formalizes semantic fidelity using a Wasserstein-distance-based criterion and demonstrates semantic-consistent channel generation through case studies that emphasize single- and multi-view sensing tasks. This work enables controllable simulation, dataset generation, and reproducible benchmarking for ISAC design and standardization, paving the way for scalable, semantics-aware channel models that bridge sensing and communication.
Abstract
Integrated sensing and communication (ISAC) increasingly exposes a gap in today's channel modeling. Efficient statistical models focus on coarse communication-centric metrics, and therefore miss the weak but critical multipath signatures for sensing, whereas deterministic models are computationally inefficient to scale for system-level ISAC evaluation. This gap calls for a unifying abstraction that can couple what the environment means for sensing with how the channel behaves for communication, namely, environmental semantics. This article clarifies the meaning and essentiality of environmental semantics in ISAC channel modeling and establishes how semantics is connected to observable channel structures across multiple semantic levels. Based on this perspective, a semantics-oriented channel modeling principle was advocated, which preserves environmental semantics while abstracting unnecessary detail to balance accuracy and complexity. Then, a generative AI-empowered semantic twin channel model (STCM) was introduced to generate a family of physically plausible channel realizations representative of a semantic condition. Case studies further show semantic consistency under challenging multi-view settings, suggesting a practical path to controllable simulation, dataset generation, and reproducible ISAC benchmarking toward future design and standardization.