Biomedical research has generated enormous volumes of data, yet the lack of structured, high-quality metadata often limits its full potential. Manual curation of metadata is labor-intensive, error-prone, and non-scalable, which hinders data integration and impedes data-driven discovery. In this paper, we propose an AI-driven framework for automatically extracting and enhancing metadata from unstructured scientific texts. Our approach leverages state-of-the-art deep learning techniques, including Convolutional Neural Networks (CNNs) and Bidirectional Long Short-Term Memory (BiLSTM) networks, to identify and predict relevant metadata elements such as experimental conditions, disease categories, and biological sample types. Comprehensive evaluations on large-scale biomedical datasets, such as the Gene Expression Omnibus (GEO), demonstrate significant improvements in metadata quality and consistency. The proposed method not only facilitates data reuse and interoperability but also supports the FAIR (Findable, Accessible, Interoperable, and Reusable) principles, thereby accelerating biomedical research and discovery.