Academician Weinan E at PAIR 2026: The Second Half of AI for Science Is Rebuilding the Research System

At the PAIR Conference 2026, Weinan E's speech was titled "Some Thoughts on the Development Strategy of AI for Science." The conference was held in Shanghai from September 20 to 21, themed "Frontiers of Next-Generation Physical Intelligence," organized by the Shanghai Institute for Physical Intelligence and Robotics. He himself spoke on the morning of September 21 at the "Next-Generation Scientific Discovery" special forum.

It is worth noting his angle. According to the report's summary, he did not stop at model capabilities, automated laboratories, or specific research cases, but extended the question to research infrastructure and the research organization system: if AI for Science truly becomes a new paradigm, what kind of infrastructure needs to be built? What different technological paths are China and the U.S. forming? When AI greatly lowers the threshold for computation, experimentation, and knowledge acquisition, what changes will occur in the existing research system?

His framework divides AI for Science into two levels. The first level is that more and more research and development activities are moving toward "industrialization," requiring a new set of infrastructure to support large-scale research activities; the second level is that productivity changes further drive changes in research "production relations," including breaking boundaries between disciplines, between theory and experiment, and between research and industry. The latter is what he calls the "second half."

In Weinan E's description, the technological paths for promoting AI for Science in China and the U.S. are not entirely the same. He said the U.S. previously relied more on landmark key projects, such as AlphaFold and materials structure prediction; these works are important, but they are more breakthroughs on a specific scientific problem, and at the time no infrastructure system covering the entire research process had been formed.

He said the turning point for the U.S. was the "Genesis Plan" proposed by the White House last year, the first time the U.S. proposed building new research infrastructure to drive an overall research paradigm change. At the same time, companies like OpenAI and Anthropic are also increasing investment in the scientific field, with a path of relying on general foundation models to continuously improve scientific research capabilities, then combining with automated, intelligent experimental systems. It should be noted that this information about U.S. policy and companies comes from the speaker's account, and the report did not provide first-hand documents.

On the Chinese side, he said that since 2022, China has proposed and promoted the construction of new research infrastructure, with four core parts: computing power; models, algorithms, and software; databases and knowledge bases; and automated, intelligent laboratories. Doing research on top of this infrastructure, he calls it "platform research"—scientists entering a specific field feel like developers developing mobile apps on the Android platform, and teams do not have to rebuild basic capabilities every time.

While automated laboratories are receiving widespread attention, Weinan E reminds us of another easily overlooked aspect: theory and computation. His reasoning is that automated laboratories solve the automation of execution and operation, but it is difficult for them to directly generate enough new scientific hypotheses; data already in the literature is always limited. The ability to truly help people break through cognitive blind spots and propose new questions and hypotheses ultimately still comes mainly from theory and computation.

He took materials and chemistry as examples. One of the most basic computational tools is molecular dynamics, which describes interactions between atoms. He said his team started working on neural network-expressed molecular dynamics potential functions very early, and later Wang Han and others further proposed the large atomic model (DPA), hoping to build a general molecular dynamics model, and this work has persisted since 2022. He said DPA-4 can already be applied to different scenarios such as organic molecules, catalytic reactions, and complex interfaces, and the next goal is to truly move toward materials design.

But he also gave a clear boundary: pushing computation further into materials process design is currently not possible. The time and space scales involved in processes are much larger. He said that although the computational throughput of molecular dynamics has increased by several orders of magnitude in the past few years, there is still a gap of two to three orders of magnitude from truly completing material composition and process design from first principles. Only when such computational tools truly mature can new materials and new molecules be designed like designing a house, and this has not yet been truly achieved globally.

Weinan E said China has already made many arrangements in automated and intelligent laboratories, including Beijing, Hefei, Suzhou, Xiamen, Shenzhen, Yibin, and other places, and Shanghai also has some key projects underway; from the perspective of overall scale and system maturity, there is still room for development, but this is undoubtedly a very important investment direction for the next three to five years.

He pointed out a basic technical bottleneck: how to automatically characterize experimental results. Taking nuclear magnetic resonance in molecular and materials research as an example, for decades this has almost entirely relied on the personal experience of senior scientists to read NMR spectra and infer molecular structures. If a truly automated laboratory is to be built, these steps must be algorithmized.

The example he gave is the SpecXMaster platform, a collaboration between Professor Zhu Rong's team at Peking University and DP Technology. He said the platform went online in March this year, has already been used by several thousand chemists, accumulates tens of thousands of feedback data entries per month, and its parsing capability on related tasks has reached a high level, at least close to that of an excellent doctoral student. These numbers and judgments are all the speaker's own account, and the report did not provide third-party evaluations or benchmark data.

He also used this case to observe the difference between general models and specialized scientific models. He said Claude had previously demonstrated the ability to read NMR spectra: if only pure NMR data and molecular formulas are provided, the success rate is actually very low; but if other information such as reactants is also provided, the recognition ability improves significantly. In his view, the characteristic of general models is that the richer the information and the more usable context, the better the overall performance; while in scenarios where input information is limited and inference must rely purely on specialized physical and chemical laws, specialized models trained specifically for scientific tasks still have a clear advantage.

Extending from this, Weinan E believes that a core hidden danger for China's AI for Science at present is still the capability of the underlying foundation model, especially when facing the most difficult tasks, where there is still a gap with the most advanced international models. He mentioned that earlier this year a Peking University team tested domestic and foreign models on the most optimized scientific research tasks, and at that time the gap between domestic and foreign foundation models was still quite obvious; although the overall capability of domestic foundation models has improved significantly now, the gap still exists on tasks that reach the top level of difficulty.

He said his team spent a year last year developing the scientific foundation model Innovator and released it at the World Artificial Intelligence Conference. But after completing this work, he felt more and more deeply: if one simply takes an existing open-source general model as the core and fine-tunes it or stacks various scientific capabilities on top, this kind of "scientific foundation model" may to some extent be a false proposition, because it does not solve the core difficulty.

He also gave a qualification: this is not to say that scientific models should not be developed, but to emphasize that if the underlying core itself is not strong enough, merely stacking scientific data on top will still severely limit the ceiling when facing the deepest scientific tasks. He also admitted that this approach is not meaningless—it can achieve good results on general scientific problems—but those who truly do frontier research often care most about the hardest part of the problems, and it is precisely on these hard bones that the gap in the underlying reasoning core is amplified most obviously.

As new research productivity is gradually established, Weinan E believes that the "second half" that follows is a change in the research system itself, with at least "three breaks": breaking the boundary between theory and experiment, breaking the boundary between disciplines, and breaking the boundary between research and industry.

The first break targets the distance between theory, computation, and experiment. He used his own experience to illustrate: in the past, he often encountered situations where a beautiful computational method was designed and the paper received high academic recognition, but there was still a big distance from the actual experimental scenario. In traditional engineering fields such as bridges, buildings, and aircraft, this boundary has long been broken, and most mechanical verification can be completed entirely by simulation. But in materials science, chemistry, and life sciences, when facing a new drug, a new molecule, or a new catalyst, today it is still impossible to complete the design task reliably enough by computation alone. He said algorithmic capabilities have improved by several orders of magnitude in the past few years, but there is still a gap of two to three orders of magnitude from truly guiding experiments and processes, and he hopes to continue narrowing it after another two or three years of effort.

The second break targets disciplinary barriers. He took materials science as an example: in China there are different research traditions from condensed matter physics, solid mechanics, and more orthodox materials science, and the research objects are actually all materials, only the time and space scales and specific problems of concern are different. Traditional materials science relies more on experiments, experience, and data, focusing on defects, dislocations, and doping; condensed matter physics starts more from quantum mechanics, electronic band structures, and phonon spectra; solid mechanics focuses on elastic deformation and fracture of macroscopic devices. He believes that from the bottom up, science is originally unified, and the real difference between disciplines is often only the time and space scales they cover.

The third break targets research and industry. He took transistors and lithium batteries as examples: both basic inventions won Nobel Prizes, but both took decades from original innovation to large-scale industrialization. The lithium battery prototype was discovered in a foreign laboratory, while today the world's largest industrial chain is in China. He believes the reason the cycle is so long is that between a principle breakthrough and industrial implementation lies a massive amount of engineering practice and optimization, and this work itself often has extremely high research difficulty. Now, with the intervention of various AI agents, including literature review, data processing, and paper writing, AI can already complete a large number of research tasks quite well, and many intermediate steps are also making breakthroughs.

In Weinan E's judgment, a very important change is the "equalization" of research. He predicts that in the next 5 to 10 years, the threshold for research is expected to drop significantly, and innovators will no longer be completely monopolized by traditional universities and a few top laboratories; an individual or small team with an idea, with the help of public infrastructure, is fully capable of realizing their ideas. Along with this equalization, research funding sources will also change: in addition to government and national system support, industrial capital and market-oriented foundations will play a greater role. He mentioned that the Shanghai Future Industry Fund and others are already exploring this area, and at the same time, transformation facilities such as intellectual property trading platforms and pilot testing bases also need to be established at an accelerated pace.

He also criticized the existing research organization from an industrial perspective: the global research system is to a large extent a "planned economy" system led by administration and academia, where the "production" and "application" of research results are disconnected from each other, researchers are assessed mainly on publishing papers, and industry needs to fish for needles in a sea of papers. He mentioned China's experience in setting up "special zones" in past economic reforms, and believes that the research system will similarly need similar attempts in the future—directly promoting comprehensive reshaping within the existing system faces great resistance, so it is possible to explore the establishment of new research institutions and public research platforms, driven by solving actual industrial problems.

Finally, he emphasized the mechanism for disruptive innovation. He believes that no matter how high AI pushes the efficiency of research engineering, free exploration based on human curiosity is always the most core source of disruptive scientific and technological achievements, and must be protected and supported; but at the same time, it is also necessary to rethink what truly counts as "free exploration"—if merely writing an ordinary academic paper is equated with free exploration, the threshold is obviously set too low.

He also gave several personal judgments: mathematics may experience a period of "confusion," because artificial intelligence's continuously improving capabilities in theorem proving, symbolic reasoning, and complex mathematical problems will directly impact the working habits and methods of mathematical research; the mathematical and physical foundations and basic principles behind artificial intelligence will become a very important mainstream direction in the next stage, and models cannot forever be made into black boxes whose workings are known but whose reasons are not; in the coming years, the integration of research resources will become increasingly important, and universities and research institutes may not need every team to repeatedly build a complete set of laboratories, but should move more toward co-construction and sharing of public infrastructure. He also judged that the past single evaluation mechanism centered solely on papers will be gradually broken, and "whether it can truly create value" will become an increasingly important yardstick.