Inside China’s “AI for Science” Strategy

At its developer conference last week, Google unveiled Gemini for Science, a new suite of tools designed to help scientists design and run experiments.

When Demis Hassabis and John Jumper of Google DeepMind won the Nobel Prize in chemistry for AlphaFold2, an artificial intelligence (AI) system for predicting protein structure, it was in recognition of a specific, bounded scientific function. The company, along with its peers, now appears to be expanding its vision of AI’s scientific deployment to include soon-to-be autonomous agentic scientists that humans will work alongside, or even for.

Google’s Pushmeet Kohli last week published a journal article representative of this more ambitious goal. In it, he describes Google’s test of AI for science tools in biomedicine along with collaborators: “The results were promising, suggesting that AI will help shape the research agenda itself by surfacing salient questions before anyone thinks to ask them,” he wrote. “Future systems will be increasingly general…We are moving toward AI that doesn’t just facilitate science but begins to do science.”

Washington hopes to support these loftier goals for AI’s scientific integration. The Donald Trump administration in November 2025 launched the Genesis Mission, a multi-departmental effort to bolster American researchers’ use of AI tools for scientific discovery in areas like biotechnology, critical materials, fusion energy, quantum, and semiconductors. The official launch video explains that the mission will unite the country’s “brightest minds, most powerful computers, and vast scientific data into one living system for discovery.” The mission, the narrator explains, will “radically redefine the scale, speed, and purpose of scientific discovery in America.”

These efforts are laudable, but they also come with a host of unanswered questions. The benefit of using an AI system to fold proteins or, if we’re lucky, cure cancer is clear. But promoting AI tools to play the role of scientist itself means the direction and output of scientific experimentation will become increasingly opaque. And U.S. labs are not alone as they reckon in real time with how both targeted and general AI tools will affect science. For over half a decade, Chinese institutions have been exploring ways to use AI to enhance research under the official term AI for Science (AI4S or 科学智能).

China’s first AI4S-specific research institution was established in Beijing in 2021. Since then, the AI4S effort has only gained steam across China’s research and policy communities as Beijing’s national strategy for scientific breakthroughs, like Washington’s, increasingly relies on AI.

AI4S means deploying AI tools throughout the scientific process. It is used for generating research questions, hypothesis formation, experimental design, data analysis, and simulation. It is a global uncertainty if, how, and to what degree the use of AI tools for scientific discovery will be institutionally or formally regulated. Automating steps in the scientific process may appeal to scientists because it can increase output and alleviate obstacles for non-native English speakers, for example. But it does not necessarily improve research quality.

A study from January 2026, notably co-authored by a team from the University of Chicago and Beijing National Research Center for Information Science and Technology, found that heavy use of AI tools in natural science research resulted in less intellectually creative papers. Such research was also tilted in favor of areas that already had large existent datasets and led to fewer follow-on studies than non-AI-assisted research. In other words, AI can help scientists publish quickly and at high volumes, but if it is over-relied on, it can also restrict the novelty and utility of the research.

Still, results so far show China’s early efforts on AI4S are paying off, even as the ultimate trade-offs merit further scrutiny. A study published last year by the European Commission found that “China has taken the lead in AI-driven research, outpacing both the US and the EU, not just in sheer output, but also in terms of scientific novelty and impact.”

In China, realizing the potential of AI4S is both a grassroots goal—research scientists hope to turbo-charge their work through effective AI assistance—and a strategic priority for top leaders. In February 2025, Gao Wen, director of the state-backed Pengchang AI Lab, published a commentary in People’s Daily echoing Xi Jinping’s description of AI as a strategic technology that will “lead this round of scientific and technological revolution.”

U.S. counterparts have only recently started speaking about AI for Science in similar terms: if Washington’s policy response loses steam, China could leverage its already substantial progress on AI4S to win the strategic advantage in one of the clearest applications of AI’s promise.

Major Chinese Policy Support for AI4S

China has two unique AI4S advantages: (1) state support for AI for Science, including resource and political commitment; and (2) scientist E Weinan, who coined the term “AI4S” in 2018. E is a Chinese Academy of Sciences academician who earned his doctorate at the University of California, Los Angeles, and taught at Princeton for 25 years. He has significantly contributed to China’s efforts to champion the AI4S concept and application.

China’s approach has been largely scientist-led, but then bureaucratized by various government entities and scientists looking to contribute to the banner national AI4S policy. China’s AI4S-related policies are under the purview of the Ministry of Science and Technology (MOST). In March 2023, MOST and the National Natural Science Foundation jointly formally launched the AI-Driven Scientific Research Initiative (人工智能驱动的科学研究转项).

During the announcement, MOST officials declared that “AI for Science has now become a new frontier of the global AI sector.” In an expert interpretation of the AI4S plan, director of the Chinese Academy of Science’s Institute of Automation Xu Bo said that a “cutting-edge AI for Science R&D ecosystem” is critical for “enhancing the competitiveness of basic scientific research.”

Partly to bolster basic research, MOST’s leadership over AI4S helps ensure access to AI resources concentrated in the technology sector can be shared with China’s research institutions. That rebalancing is in line with the Natural Science Foundation’s (NSF’s) stated mission of funding projects, including those that fall under the AI4S umbrella, that possess a “strong exploratory nature” and feature “novel topics.” Specifically, a request for proposals issued by the NSF mentioned interest in funding projects that develop foundational AI algorithms for science-focused models.

AI4S has also been explicitly embedded in China’s national supercomputing network: in April 2016, Sugon, a major hardware producer with strong ties to the Chinese Academy of Sciences (中国科学院), announced the launch of a massive AI4S-specific computing cluster. The facility will be located in central China’s Henan province in Zhengzhou. At the press conference for the site, Sugon Senior Vice President Li Bin put it plainly: “In the new era, supercomputing must be geared toward AI4S.”

Developing LLMs for Science

In parallel to generalized AI4S progress, Chinese research institutions and firms are working to create large language models (LLMs) suited for specific scientific purposes. For example, the Baiyulan (白玉兰) series of LLMs was developed in 2023 by the Shanghai Baiyulan Open Source Research Institute, a joint venture lab led by Shanghai Jiaotong University.

Baiyulan released five different domain versions of its sci-LLM, including:

• Baiyulan-Benchmark (SCiEval): trained on scientific data across major scientific fields and meant to act as an evaluator for the scientific capabilities of other LLMs, testing them in basic knowledge, knowledge application, and research capabilities
• Baiyulan Chemical Synthesis (BAI-Chem): specialized in chemical synthesis research
• Baiyulan Chemical Synthesis 2.0 (BAI-Chem 2.0): a more powerful iteration of the original
• Baiyulan-Neurofluid (BAI-NeuroFluid): specialized in using physics-inspired generative AI to simulate fluid particle models and dynamics

But even before specific science models are rolled out, foundational models could be sufficient to propel China’s AI4S strategy. Research published last year showed DeepSeek consistently achieved either the highest or second-highest accuracy score on college and graduate-level materials science questions.

DeepSeek-R1 reportedly beats OpenAI’s o1 on certain math and science-related benchmarks. Preliminary tests of R1’s performance on data-driven scientific tasks, taken from real papers in topics like bioinformatics, computational chemistry, and cognitive neuroscience, show it matched o1, according to Sun Huan, an AI researcher at Ohio State University. Sun’s team also found running scientific inquiries in R1 cost 13 times less than o1.

It is still too early for retrained or fine-tuned DeepSeek4Science LLMs to have fully emerged in the literature. But it is likely only a matter of time: DeepSeek LLMs specifically for AI4S applications are almost certainly forthcoming.

AI for Nuclear Fusion in China

China has created a research center dedicated to AI4Fusion, a subspeciality under the broader AI4S banner. It’s called the Hefei Comprehensive National Science Center (Institute of Energy) Artificial Intelligence and Digital Twin Research Center (合肥综合性国家科学中心 (能源研究院) 人工智能与数字孪生研究中心). According to the center’s website, it is focused on “leveraging AI and large-scale modeling technology to empower fusion reactor plasma control and simulation.” Its research focus areas include working on digital twin systems for tokamaks, AI-enabled fusion device control, and commercial applications of multi-process real-time control and data acquisition technologies.

Endeavors like the Hefei AI4Fusion center are testament to Chinese scientists’ recognition of what’s at stake. According to Sun Xuan, the founder of Chinese fusion energy startup Xeonova and a professor at the University of Science and Technology of China’s School of Nuclear Science and Technology, breakthroughs in either nuclear fusion or AI individually could “represent the arrival of a new era.” But if the two can work together, he continued, “it may accelerate the arrival of this great era.”

Over the last two years, the new China National Nuclear Corporation (CNNC) subsidiary China Fusion Energy Co., Ltd. and other state-backed enterprises have received roughly $4 billion in funding. New private Chinese fusion startups such as Xeonova reportedly raised $2 billion in investment last year.

AI tools have become a prominent topic in China’s fusion research community. An April 2025 article in China Energy News reposted by the People’s Daily stated that researchers at CNNC’s Southwest Institute of Physics in Chengdu utilized AI-enhanced simulations to reduce the preparation time per plasma shot at the institute’s Circulator-3 (aka Huanliu-3) experimental tokamak from several days to just half an hour.

Support for AI for fusion has reached the national level. In October 2025, the Communist Party of China Central Committee released a set of recommendations on the 15th five-year plan, which stated that China needs to “explore diverse technology roadmaps, typical application scenarios, feasible business models, and market regulation rules and work to foster new drivers of economic growth,” including nuclear fusion and AI.

The National Key Project for Research and Development of Magnetic Confinement Nuclear Fusion Energy is another longstanding MOST-led fusion initiative with clear relevance to AI4S. The program has already been budgeted over $160 million for research projects covering how AI can enable intelligent control of plasma pressure profiles in tokamaks, real-time plasma pressure control algorithms, and remote-controlled vacuum leak detection for fusion reactors, among other subtopics.

It is likely that forthcoming local, industrial, or ministerial-level implementation plans will explicitly advocate for AI-driven fusion research in support of China’s broader AI4S strategy.

How AI is Advancing Materials Science in China

Before the dawn of AI4S, China was already doubling down on materials science research. China replaced the United States as the top producer of materials science output in 2018 and is also the fastest-rising country in the field, according to the Nature Index. From 2008 to 2019, China’s National Natural Science Foundation quadrupled its funding for engineering and materials science, from around $70 million to around $280 million.

China’s success in materials science, then, gave it a strong leg up as the country began experimenting with AI for (materials) Science, or AI4(M)S. These efforts benefit from the dual state goals of upgrading the industrial base and achieving the lead in critical materials research.

Using AI4(M)S has been a focus of both government leaders and key figures in the field. It is consistently touted by the prominent academician Xie Jianxin, who has expressed the view that AI4(M)S is essential for gaining the strategic advantage in sci-tech and innovation. He has credited the United States’ Materials Genome Initiative as well as what he called “the recent paradigm shift in materials science research and development driven by data and artificial intelligence” for attracting unprecedented attention to the field. His vision is to help China harness AI’s potential to lead the world in the field.

Central-level policymakers are taking proactive measures to make that vision a reality. In 2022, MOST and five other ministries released opinions on promoting “high-quality” economic development through AI, specifically calling for “high-level scientific research in materials science and new materials.”

One of the most notable recent policy developments was the 2024 New Materials Big Data Center Overall Construction Plan (新村料大数据中心总体建设方案), jointly issued by the Ministry of Industry and Information Technology, Ministry of Finance, and the National Data Administration. It provides the most comprehensive level of detail on how AI4S will be implemented in the materials science domain, including constructing a national data center in Beijing and an online platform aggregating materials science literature, datasets, and functional AI models. The plan states:

Data, as a new production factor, is a strategic resource critical for new materials technological innovation and high-quality industrial development…Through the materials + data model, we will facilitate original innovation in new materials, serve the development of new materials enterprises, cultivate new engines for innovation and development in the new materials industry, accelerate the formation of new-quality productivity, and build new global competitive advantages.

So far, China’s AI4S strategy has helped AI-enabled materials science research institutes make significant advancements. The Beijing Advanced Innovation Center for Materials Genome Engineering created a specialized LLM, called SteelBERT, to predict the mechanical properties of steel alloys based on fabrication processes entered as sequences of descriptive text.

The private sector has also benefitted. Companies like DeepChem are developing a host of AI tools for materials science research, such as autonomous “intelligent laboratories” that can be programmed to carry out high-throughput experiments 24 hours a day. On the more mainstream commercial side, Xiaomi has already reportedly used AI tools to design novel lightweight metal alloys for next generation EVs.

DPTech’s large-scale AI models have purportedly helped optimize the production of sodium-ion cathode materials. The company claims it has an AI-enabled proprietary platform called “Piloteye” that models the entire production lifecycle of battery materials. DPTech also reportedly cooperated with the Beijing Institute of Scientific and Intelligent Research (北京科学智能研究院) to launch the OpenLAM Large Atom Model Project, iterating on a pretrained potential energy model called DPA-2 which can predict the behavior of over 90 elements in the periodic table.

Co-founder of DPTech and president of the AI for Science Institute in Beijing Zhang Linfeng cited battery materials as a key area where AI tools can help drive not just scientific breakthroughs, but also commercial applications.

Zhang’s team has reportedly worked with battery giant CATL to develop techniques that overcome the enigmatic growth of lithium dentries using specialized platforms. Zhang was one of a handful of tech leaders Xi met with during his inspection tour of Yizhuang, a tech hub in Beijing.

China’s foray into AI for materials science stands to accelerate progress in key industrial applications, possibly reducing reliance on foreign inputs, as well as the country’s military and space programs. Materials science offers an illustrative case study into how Beijing is strategically promoting AI tools to expedite progress in a domain China already leads in.

AI for Science or AI Instead of Scientists?

“Will AI replace scientists?” Chengzhi Wang, CEO of biotech company Zhiyuan Shenlan, wrote in an article discussing the future of AI and science in China (translated by Jeff Ding). Wang continued: “My answer is straightforward: it will eliminate mediocre scientists, but it will empower top scientists… Currently, AI still needs human guidance, but at the execution level, it is about to deliver a devastating blow.”

Dongzhan Zhou, a scientist at the AI for Science Center, Shanghai Artificial Intelligence Laboratory, described building an agent for the purpose of scientific research. But Zhou hopes that the relationship between scientists and AI does not become “one of replacement.” The foundational value of human scientists, Zhou writes, “lies in discovering new scientific problems and defining those problems.” In other words, AI can help move the research mechanics forward, but it should not be the one setting the research question or parameters.

Similar anxieties have infiltrated the American scientific community. Dr. Brian Keith Spears, a fusion scientist at Lawrence Livermore National Laboratory who experimented with GPT-5 in his research, described the model as useful yet imperfect. In its eagerness to please, he said, GPT “often introduces numerical duct tape to smooth over a thorny issue, silently swaps out detailed numerical solves for approximations with trends it knows I want, and confidently declares victory when numerical signals are still obviously noise.”

A Nature study published in January 2026 highlighted another potential limitation. Since AI systems are most useful when working through large datasets, widespread use of AI4S could bias research toward subjects where such troves of data already exist. That stands to constrain the frontier, disincentivizing exploration of wholly novel subjects.

The distinguishing factor for Chinese researchers, however, is that their experimentation with AI is largely supported and subsidized by the government. Officials will be paying close attention to if and how AI tools help China make progress on its national sci-tech objectives, and they can intervene in the event of seismic employment effects on the innovation talent base. What stands now as a two-way street between the scientific community and government officials could pay off in the long term, if state leaders listen to the insights and priorities of researchers on the ground.

Jimmy Goodrich is an IGCC senior fellow. The author would like to thank China technology writer Johanna M. Costigan and Mykael Goodsell-Sootho, a China analyst, for significant research and writing contributions to this piece.

Thumbnail credit: Josh Chin (Flickr)