Vision Language Models for Radiation Patterns to Antenna Parameters

Abstract

Observed radiation patterns often serve as a primary evidence of antenna's behavior, but translating them into meaningful interpretations is a nontrivial and expertise intensive task. This demand necessitates automated pattern interpretation, a diagnosis problem encountered in applications encompassing Radio Frequency (RF) surveillance, non cooperative emitter characterization and Over The Air (OTA) testing. This work addresses the incorporation of Contrastive Language Image Pre training (CLIP) and other recent vision language models to perform such a diagnosis. These models analyze the multimodal data and extract discriminative features directly from radiation pattern images, to train Machine Learning (ML) approaches for meaningful inferences on geometrical and performance parameters of the antenna. Performance of three such vision language models is demonstrated using RadPat 50K, a synthetic dataset of radiation pattern images generated for uniform linear arrays (ULAs). The radiation pattern images are processed by the vision language models to obtain discriminative features. The CLIP extracted features are processed by either ML classification models to infer antenna parameters number of elements, element spacing, weighting scheme, presence of grating lobe and steering angle, or ML regression models to infer parameters beam width, directivity and main lobe direction. ML models achieve accuracies exceeding 70 percent, highlighting the potential of multimodal Artificial Intelligence (AI) towards intuitive antenna analysis.