Channel Knowledge Map Construction via Physics-Inspired Diffusion Model Without Prior Observations

The ability to construct channel knowledge map (CKM) with high precision is essential for environment awareness in 6G wireless systems. However, most existing CKM construction methods formulate the task as an image super-resolution or generation problem, thereby employing models originally developed for computer vision. As a result, the generated CKMs often fail to capture the underlying physical characteristics of wireless propagation. In this paper, considering that acquiring channel observations incurs non-negligible time and cost, we focus on constructing CKM for large-scale fading scenarios without relying on prior observations, and we design three physics-based constraints to characterize the spatial distribution patterns of large-scale fading. By integrating these physical constraints with state-of-the-art diffusion model that possesses superior generative capability, a physics-inspired diffusion model for CKM construction is proposed. Following this motivation, we derive the loss function of the diffusion model augmented with physics-based constraint terms and further design the training and generation framework for the proposed physics-inspired CKM generation diffusion model. Extensive experiments show that our approach outperforms all existing methods in terms of construction accuracy. Moreover, the proposed model provides a unified and effective framework with strong potential for generating diverse, accurate, and physically consistent CKM.

💡 Research Summary

This paper proposes a novel physics-inspired diffusion model for constructing Channel Knowledge Maps (CKMs), specifically Channel Gain Maps (CGMs), which are crucial for environment-aware 6G wireless networks. The core challenge addressed is generating accurate, physically consistent CKMs without relying on costly and time-consuming prior channel measurements, using only readily available inputs like building maps and transmitter locations.

The authors identify a key limitation in existing CKM construction methods: most leverage models (e.g., GANs, diffusion models) directly adapted from computer vision, which prioritize visual fidelity over adherence to the underlying physics of wireless signal propagation. This leads to generated maps that lack physical plausibility.

To bridge this gap, the paper introduces three heuristic physics-based regularization terms designed to capture the macroscopic spatial distribution patterns of large-scale fading: 1) Edge Loss: Penalizes discrepancies in signal strength gradients at building boundaries, reflecting abrupt attenuation near obstacles. 2) Regional Propagation Loss: Encourages consistent signal decay patterns within homogeneous regions (e.g., open areas, building clusters). 3) Multi-scale Feature Loss: Ensures the generated map exhibits correct channel characteristics across different spatial scales.

The main technical contribution is the seamless integration of these physical constraints into a state-of-the-art denoising diffusion probabilistic model (DDPM). The authors derive a modified training objective function that augments the standard diffusion model’s evidence lower bound (ELBO) loss with the proposed physics-based regularization terms. This forces the model to learn not only the statistical distribution of possible CGMs but also the fundamental laws governing radio wave propagation. Furthermore, a tailored three-stage training framework is designed to efficiently compute these physics terms during optimization.

Extensive experiments demonstrate that the proposed model significantly outperforms all existing baselines, including spatial interpolation methods (Kriging, IDW), deep learning models (UNet, cGAN), and standard conditional diffusion models, in terms of construction accuracy (e.g., lower MAE and RMSE). The generated CGMs show sharp transitions at building edges, smooth decay in open areas, and high physical consistency, especially in complex urban canyon scenarios. Ablation studies confirm the individual contribution of each physics term to the final performance.

In conclusion, this work presents a unified and effective framework for generating diverse, accurate, and physically faithful CKMs. It represents a successful application of physics-informed machine learning to wireless communications, offering a powerful digital twin tool for 6G applications like UAV trajectory planning, BS deployment, and network resource optimization.