Recent years have known the development of radio propagation models (RPM) and specially the AI-based ones. These models are interesting for many applications such as radio planing and design, fingerprinting-based localization, radio resources management etc. However most of the proposed AIbased RPMs have been trained on simulated radio data making them not ready or reliable for real condition applications. In this work we tackle the problem of learning and fine-tuning an AI-based RPM from simulated data to real world radio measurements. We raise the inherent problems and limitations and then propose solutions to overcome them. The study has been focused on 5 GHz Wi-Fi and Home network environment.