BEIJING, July 26 (Xinhua) – The Chinese Academy of Sciences (CAS) unveiled science models on Saturday at the 2025 World Artificial Intelligence Conference, showing significant advances in AI-driven research innovation.
ScienceOne is designed to understand complex scientific data such as waveforms, spectra, and fields, and integrates the capabilities of literature extraction, knowledge inference and tool orchestration, aimed at serving as an intelligent foundation for innovation in a variety of fields.
AI offers an unprecedented opportunity to transform basic research methodologies, accelerate scientific discovery and address humanity's most pressing scientific challenges. However, current approaches rely primarily on generic models that use domain-specific data to fine-tune them while developing isolated, single-field tools. This paradigm faces three important constraints: It is a fragmented scientific data system, insufficient professional reasoning ability, and a closed research ecosystem.
To overcome these challenges, 12 CAS labs have collaborated on and developed ScienceOne, including the Automation Laboratory, the Center for Computer Network Information, the National Library of Science, the Academy of Mathematics and Systems Sciences, and the Institute of High Energy Physics.
The model provides a comprehensive understanding of the modalities of complex scientific data such as waveforms, spectra, and fields, while integrating the core features of literature extraction, knowledge inference, and tuning computational tools. Researchers have easy access to model functions at various stages of scientific research, allowing for seamless integration of AI into scientific research.
ScienceOne deeply customizes scientific customization based on the China Open Source Foundation model, thus integrating a suite of unique models developed for common scientific data types and incorporating domain-specialized tools such as Alphafold and Matthgen.
It demonstrates the systematic acquisition of core principles, law, and specialized knowledge across fundamental fields such as mathematics, physics, chemistry, astronomy, geoscience, and biology.
Evaluations confirm cutting edge performance in mathematics, physics, chemistry, materials science and biology, with key capabilities in calling and inference of scientific tools. ScienceOne achieved even more excellent results in the final exam of humanity (HLE).
Based on ScienceOne, the research team also developed two scientific intelligent agents.
One is designed to help researchers read articles thoroughly, write reviews, and evaluate research topics and technical pathways. Get access to 170 million scientific literature and real-time open source science information. With that support, a literature review task that previously required 3-5 days can be completed in just 20 minutes.
Other agents aim to lower barriers to the use of scientific tools, allowing for autonomous planning for over 300 scientific computing tools. It can automatically identify research tasks and intelligently schedule the best tools, thereby increasing the efficiency of the research process.
ScienceOne now drives X-cell digital cell platforms that accelerate discovery across fields, automate biological target identification, improve particle simulation efficiency in Beijing Electronic Positron Collider, improve molecular prediction accuracy in chemistry, streamline global stretch coordination for astronomy, and support structural design innovations of fast design innovation. ■
