THE Chinese Academy of Sciences (CAS) has launched a large collection of curated scientific and technological texts and data designed to help artificial intelligence models understand specialized scientific knowledge.

Named the AoCang S&T Corpus, the resource draws on scientific data, research papers, invention patents and sci-tech books that have been processed into material AI systems can use.

Led by the National Science Library of the CAS, this project consists of one national database integrating knowledge across disciplines, seven specialized databases and multiple datasets designed for specific applications. It covers fields including mathematics, physics, chemistry, chemical engineering, astronomy and space science.

Qu Jiansheng, director of the library, said the corpus could help fill gaps in AI models’ specialized knowledge and correct errors in logical reasoning, making their outputs more consistent with basic scientific principles and mathematical logic.

Beyond research, the project will provide tailored datasets for commercial spaceflight, intelligent manufacturing and biomedicine. It is expected to support more than 100 industry applications, such as electronic component selection, ultrastructural pathology diagnosis and high-speed train design.

The AoCang Corpus distinguishes itself from generic internet-scraped training data through its rigorous curation and quality-control pipeline. Unlike raw web text, which often contains misinformation, ambiguities, or superficial explanations, the corpus undergoes expert annotation and semantic structuring to preserve the intrinsic logic of scientific discourse. This process ensures that the underlying data not only provides factual answers but also embeds the procedural reasoning and experimental context necessary for complex problem-solving. By feeding AI systems with this high-fidelity knowledge graph, researchers aim to move beyond simple question-answering toward genuine scientific inference, enabling models to generate hypotheses, verify calculations, and trace the evolution of theoretical frameworks across decades of published research.

(Xinhua, SD-Agencies)