China launches national scientific corpus to support AI-driven research

China on Friday unveiled a national scientific corpus infrastructure at the 2026 World Artificial Intelligence Conference in Shanghai, aiming to provide the high-quality data foundation needed to support the next generation of AI-powered scientific research.

The Aocang S&T Corpus, developed by the National Science Library of the Chinese Academy of Sciences together with nearly 100 institutions, is designed to serve as a national infrastructure of high-quality, domain-specific and AI-ready data covering all major scientific disciplines.

Named after Aocang, the largest state granary of the Qin Dynasty (221-206 BC), which was built in present-day Henan province to supply grain for the newly unified empire, the corpus carries the same strategic logic into the digital age, said Qu Jiansheng, director of the National Science Library.

The launch comes as AI for Science — the use of artificial intelligence to accelerate scientific discovery — has become a strategic priority, with China's scientific leadership emphasizing the need to advance AI-enabled research to support the country's transition from a global contributor to a frontier innovator, he said.

Unlike general-purpose internet data, scientific data are highly structured, knowledge-intensive and often fragmented across institutions. Building trustworthy scientific corpora has become a common challenge for AI research worldwide.

Aocang is designed to address that challenge by transforming heterogeneous scientific resources into standardized, AI-ready corpora for scientific discovery.

"Aocang holds irreplaceable, foundational and strategic value for strengthening the data bedrock of national AI development and achieving high-level scientific and technological self-reliance," Qu said.

Drawing on the academy's repositories, Aocang integrates more than 320 petabytes of scientific data, 150 million research papers, 120 million patents and 5 million scientific books.

Organized into a national core library and multiple specialized databases, the corpus covers all major disciplines — from fundamental physics to industrial innovation — and supports diverse data formats, including text, formulas, spectra and waveforms.

To bridge the gap between raw data and machine-readable knowledge, developers created a three-tier refinement framework that upgrades basic datasets into feature-rich and knowledge-dense corpora. The process significantly improves semantic density, producing what Qian Li, deputy director of the National Science Library, described as "AI-ready" training material.

"These high-knowledge, high-semantic-density corpora can fill gaps in large models' domain expertise, correct logical reasoning biases, and equip them with genuine disciplinary understanding and long-term reasoning capability," Qian said.

Backed by more than 520 proprietary software tools, Aocang has established a fully automated pipeline for data processing, quality control and trusted data sharing.

The platform is designed to expand continuously through contributions from the global research community.

Preliminary results show that Aocang has enhanced the reasoning capabilities of the CAS-developed ScienceOne foundational science model.

The corpus is also being integrated into major commercial models, including Doubao, Qwen and Ant Group's medical AI model.

In its first phase, the project aims to integrate more than 36 billion scientific resource entries across 2,883 high-quality scientific corpora and 68 petabytes of research data, supporting more than 500 application scenarios.

Qu said the academy will continue expanding the corpus ecosystem, strengthening data-processing capabilities and exploring hybrid operating models that combine public-interest access with sustainable commercial mechanisms.

"We hope that our corpus will empower new quality productive forces and contribute a reliable, world-class data foundation to global scientific innovation," he added.