A Diffusion-Based Propagation Model for Path Loss Prediction in Indoor Environments

Machine learning methods have been widely used to design accurate and computationally efficient propagation models. Various machine learning models have been used for this task, ranging from simple artificial neural networks to more complex deep learning structures. In most cases, these models are trained utilizing either measured or simulated data. In this paper, we introduce a new approach to propagation modeling in indoor environments, leveraging a diffusion model. Our model is trained using simulated path loss maps in many diverse indoor geometries. Then, it can efficiently extrapolate full path loss maps from a small number of path loss samples, collected by measurements or low-cost simulations. This form of path loss “inpainting” requires no re-training of the model, and is agnostic to the shape and source of the provided path loss samples.