Deep learning for direction of arrival estimation: a review of hybrid model-data fusion

Traditional Direction of Arrival estimation methods degrade markedly under non-ideal conditions. Deep learning offers a robust data-driven alternative; however, the rapidly expanding method landscape calls for a structured overview. To address this need, this review first proposes a systematic classification framework for deep learning-based DOA techniques. It encompasses architectures from foundational CNNs to advanced Transformers, discussing their respective strengths in feature extraction and signal reconstruction. With a particular focus on the emerging paradigm of hybrid model-data fusion, this review analyses how methods like DA-MUSIC and SubspaceNet merge theoretical interpretability with data-driven adaptability to enhance performance. Comparative analyses indicate that these models achieve superior estimation accuracy and resolution probability, particularly at low SNR. Finally, by summarizing critical challenges and outlining future research directions in lightweight design, interpretability, and cross-scenario generalization, this paper provides a comprehensive roadmap. This review aims to guide researchers in developing the next generation of robust, efficient, and intelligent DOA estimation systems.