At the Intelligent Manufacturing Research Institute of Hefei University of Technology, staff are debugging an AI chemical management robot.

The Chinese Academy of Sciences has released the “Panshi 100” model system, which targets eight major disciplines to create a cluster of large models in various fields.

Currently, AI is unprecedentedly involved in scientific research, from predicting protein structures to discovering new materials, showcasing its potential as a “universal engine” for accelerating science.

As a new partner for researchers, how is AI changing the path and pace of scientific research? How can we use AI responsibly and effectively? This edition invites several experts to discuss these questions.

1. How is the path of scientific discovery changing?

Traditional research begins with “hypothesis-validation,” but now, the path of scientific discovery is gradually shifting towards a new paradigm of “data-pattern discovery-intelligent generation-closed-loop iteration.”

Wang Xijun, a distinguished professor at the University of Science and Technology of China, notes that in traditional research, researchers often propose questions based on experience and intuition. Now, for some disciplines, AI can proactively discover patterns in massive datasets, shifting the path of scientific discovery to a new paradigm where AI can precisely design desired materials based on target requirements.

For example, in my research on framework materials, these materials can be manufactured in massive structures through combinations of different metal nodes, organic ligands, and connection methods, reaching trillions in scale, far exceeding human exploratory limits. In this context, AI provides a breakthrough. Machine learning can quickly predict material performance, saving significant trial-and-error costs in real experiments. Furthermore, AI can extract patterns from data, transforming past intuitive approaches into computable and transferable models, making material design more rational.

On this basis, generative AI can further drive research from “selecting the known” to “creating the unknown”—directly generating new material structures beyond training data to achieve “reverse design” around target performance. This means that AI is not only accelerating problem-solving but also expanding the boundaries of the problems themselves.

Thus, AI’s role in research is continuously evolving: from an initial computational tool to a research tool that assists in analyzing patterns, and now to a “research partner” that can participate in and even drive autonomous exploration.

Of course, AI will not replace scientists. Understanding key scientific questions and mechanisms still relies on human judgment and insight. Humans are responsible for posing questions and guiding direction, while AI searches for possible answers within vast data and complex spaces. This collaboration will provide a more solid and expansive space for future research innovation.

2. Is the efficiency of research innovation improving?

AI excels at handling tasks with clear answers and requiring extensive repetitive calculations.

Mo Bofeng, a professor at the Oracle Bone Research Center of Capital Normal University, states that AI significantly enhances research efficiency in literature review, experimental design, and data analysis. Even when dealing with oracle bones over 3,000 years old, AI can play a substantial role. Previously, tasks like oracle bone stitching (reassembling broken bones) and complementing (restoring missing images) relied solely on the experience of a few experts. Now, AI provides new solutions.

To truly leverage AI, it is crucial to identify the right integration points. As oracle bones are archaeological documents, the core research goal is to restore textual materials and information, and AI is particularly adept at handling tasks with clear answers and extensive repetitive calculations. It can identify subtle features that humans may overlook, such as the curvature of fracture edges and the angles of brush strokes, providing key clues for stitching and complementing.

However, AI is not omnipotent. The total number of oracle bones exceeds 160,000, with over a million characters, which may seem substantial but is still insufficient for training AI large models. Therefore, human experts are still needed for deep semantic judgments. A more effective approach is human-machine collaboration: using AI as a speed-up tool while relying on expert judgment to review and correct its results.

Currently, stitching and complementing are just the beginning of AI-assisted oracle bone research. As technology advances, tasks like classification, aggregation, and translation of oracle bones will gradually break through. Future researchers will need not only professional knowledge but also enhanced data processing capabilities, adeptly leveraging technology to amplify their research advantages.

3. Will AI influence research judgment?

While lowering some research barriers, the risks of false citations and erroneous reasoning deserve attention.

Yang Yaodong, a researcher at Peking University’s Institute of Artificial Intelligence, explains that AI is not just helping researchers write code, review literature, and create charts; it is changing the entire research process: from a linear flow of humans posing hypotheses, conducting experiments, and analyzing results, to a closed-loop system of human-machine collaboration, model prediction, automated experiments, and feedback iteration.

This change brings several benefits. First, efficiency is greatly enhanced; in fields like materials, drugs, and energy, there are numerous candidate solutions, and traditional methods struggle to exhaust them. AI can quickly filter options, liberating researchers from repetitive trial and error to focus on key issues. Second, it promotes interdisciplinary integration, as a scientific problem often involves physics, chemistry, biology, engineering, and computation, and AI can establish connections between multi-source data. Third, it lowers some research barriers; with open-source models and tool platforms, small teams can undertake large projects.

However, it is important to note that AI does not equate to true scientific understanding. Scientific research must not only make accurate predictions but also answer “why.” If models are black boxes, data sources unclear, and experimental processes non-reproducible, the conclusions drawn by AI could pose new risks. Especially with generative AI, issues like false citations, erroneous reasoning, low-quality papers, data leaks, and unclear academic responsibilities could undermine research norms.

A deeper issue is that research judgment should not be replaced by tool logic. AI excels at finding optimal solutions within existing data, but determining what problems are worth studying and which results hold scientific significance still requires human oversight.

4. How can resources be effectively integrated?

Connecting scientists, AI engineers, and industry forces to shift innovation from isolated breakthroughs to systematic acceleration.

Wu Libo, assistant president of Fudan University and chairman of the Shanghai Institute of Science Intelligence, states that scientific intelligence is transitioning from a “technology-centered” 1.0 era to a “scientist-centered” 2.0 era. The 2.0 era aims to make more scientists the protagonists, allowing AI to truly permeate the entire research process. The Shanghai Institute of Science Intelligence and Fudan University jointly established the Xinghe Qizhi Scientific Intelligence Open Platform to respond to this shift.

The primary role of the platform is to lower the barriers for scientists to use AI. It has built a comprehensive infrastructure covering data, models, computing power, experiments, intelligent agents, and collaborative communities around real research paths. Currently, the Xinghe Qizhi Scientific Intelligence Open Platform has gathered over 400 scientific models and tools, 22PB (petabytes) of high-value data, and 500 million literature patents, allowing scientists to access cutting-edge models for research without delving into technical details.

We have also launched a research intelligent agent system represented by “Dasheng.” It can understand scientific problems and assist in completing the entire process from literature analysis, hypothesis generation to experimental validation. Recently, “Dasheng” introduced a custom laboratory function, allowing scientists to build their own toolchains based on their research directions.

The second role of the platform is to promote interdisciplinary, interregional, and cross-field integration. In traditional research, data, models, and methods from different disciplines often do not communicate, making collaboration difficult. The Xinghe Qizhi Scientific Intelligence Open Platform enables sharing, reuse, and combination of results from different fields through a unified model repository and data infrastructure.

On a deeper level, the platform serves as a hub for the scientific intelligence ecosystem. It connects scientists, AI engineers, and industry forces, allowing data and methods to flow and be reused within the system, shifting innovation from isolated breakthroughs to systematic acceleration, providing sustainable institutional support for AI-driven research paradigm transformation.

5. How to build and effectively utilize intelligent platforms?

Encouraging open sharing to bridge the gap between industry and research.

Liu Tieyan, president of Beijing Zhongguancun College and chairman of the Zhongguancun Artificial Intelligence Research Institute, emphasizes that having many platforms does not equate to them being sufficient, usable, or genuinely useful. Last year, Zhongguancun College surveyed over 30 materials companies in Beijing, identifying 100 “bottleneck” issues. The survey revealed that with current mainstream scientific intelligence technologies, only 20% of the problems are expected to be solvable. The remaining issues are temporarily unsolvable due to low levels of digitalization in enterprises, data deficiencies, and insufficient algorithm precision. This realization highlights that “AI empowering research” cannot merely be a slogan or platform; there are real challenges such as infrastructure deficits, technological limitations, and gaps between industry and research.

Regarding the open sharing of scientific intelligent agents and tools, while it may seem like a technical issue on the surface, at a deeper level, it is a lack of motivation to connect. Why would an organization open its data and platform? If this question lacks a systematic answer, “open sharing” will remain at the level of advocacy.

To break this deadlock, it is suggested to approach it from three aspects: first, vigorously promote industrial digitalization, allowing genuine industry needs to guide scientific research directions. Research should not remain in a “research first, then transform” model; industry feedback should enter the research cycle to fill the “last mile.” Second, establish incentive mechanisms for open sharing, making sharing a recognized research contribution, such as being a condition for project initiation and completion, and creating a citation metric system similar to that of papers. Third, public powers should take the lead in building the underlying infrastructure for interdisciplinary collaboration. Users of scientific intelligent agents and tools are highly specialized and dispersed across various disciplines. Due to insufficient market size, national strategic investment may be considered first, gradually introducing market mechanisms.

In summary, breaking through data and intelligent agent interfaces is a surface issue; reconstructing incentive mechanisms is a mid-level issue; and fundamentally, research must face national needs and genuine industry problems.