AI for scientific research has moved past the conceptual stage.
In 2024, the Nobel Prize in Chemistry was awarded to Demis Hassabis and John Jumper, two artificial intelligence scientists from Google DeepMind. The AlphaFold system they developed predicted the 3D structures of nearly 200 million proteins, an amount equivalent to 1,000 times the cumulative experimental results of scientists worldwide over more than 50 years.
Only by clarifying the shape of proteins can we understand diseases and design drugs. Approximately one-third of all marketed drugs worldwide target disease-related proteins. One of the prerequisites for the successful design of candidate molecules for COVID-19 mRNA vaccines within 48 hours after obtaining the viral gene sequence in 2020 was that scientists had already clarified the 3D structure of the coronavirus spike protein. The emergence of AlphaFold has significantly accelerated the efficiency of protein structure analysis, and has made AI-assisted scientific research a global focus.
AlphaFold can achieve this because the field of protein research has more than 50 years of standardized data accumulation, mature and unified detection methods, and globally shared databases, which are naturally suitable for AI modeling.
However, in many scientific research fields, the data conditions are far from being so favorable. Whether AI can play a similar role in these fields has become a common concern of both the scientific community and the industrial sector.
In August 2025, the "Opinions on Deeply Implementing the 'Artificial Intelligence +' Action" issued by the State Council explicitly proposed to "accelerate the exploration of new scientific research paradigms driven by artificial intelligence". On July 18, during the 2026 World Artificial Intelligence Conference (WAIC), four institutions - the Beijing Academy of Artificial Intelligence, the Chinese Academy of Sciences, Shenzhen Hetao College, and Tsinghua University - showcased their respective phased results at the roundtable forum titled "AI for Science: Beyond the Concept".
The four institutions are exploring different directions, but the ways AI can assist scientists have become very concrete: reading messy experimental data, operating precision instruments, and managing experimental processes. In the past, a large amount of researchers' time was spent on these repetitive tasks. By reducing the time spent on such work, scientists can focus their energy on formulating hypotheses, designing experiments, and interpreting results.
Interpreting Experimental Data
In July this year, the R&D team of the Shanghai Institute of Ceramics, Chinese Academy of Sciences identified a new type of alloy catalyst from 20 million candidate formulations, with an activity 38% higher than that of commercial catalysts on the market. A single dry experiment screening process only takes 30 minutes.
This catalyst is used in the hydrogen evolution reaction, which uses electricity to decompose water into hydrogen, and is a key material in the green hydrogen energy industry chain. Catalysts are also very common in daily life. For example, the three-way catalytic converter in car exhaust pipes is responsible for converting carbon monoxide and nitrogen oxides emitted by the engine into harmless gases; iron-based catalysts are used in fertilizer plants to synthesize ammonia, and food production for nearly half of the world's population depends on this process.
The performance of a catalyst directly determines how much energy and time a chemical reaction requires, but finding a suitable catalyst has always been one of the most time-consuming tasks in materials science. The traditional approach is to repeatedly trial and error among a large number of candidate formulations. Researchers synthesize a batch of materials, test their performance, analyze the data, and adjust the formulations. When the number of candidate formulations reaches the order of tens of millions, a full cycle often takes several months.
The tool that helped this team reduce the time from several months to 30 minutes is the ScienceOne Omni 2.0, a large scientific foundation model jointly developed by multiple research institutes organized by the Chinese Academy of Sciences. One of the core problems that ScienceOne aims to solve is that data in different disciplines are completely different in form: molecular structures are 3D coordinates, spectra are 1D waveforms, proteins are amino acid sequences, and scientific images are pixel matrices.
ScienceOne's approach is to first uniformly encode these data in different formats into a form that the model can process, then perform alignment training using 170 million scientific and technological documents from around the world and the CAS scientific corpus to enable the model to establish correlations between different disciplines, and finally output results for specific scientific research tasks. The training data contains 8 million scientific reasoning samples, and the platform integrates more than 8,000 professional scientific research tools and skill libraries.
Xu Nan, a researcher at the Institute of Automation of the Chinese Academy of Sciences, stated at the roundtable forum that what ScienceOne 2.0 aims to do is to "let the model think like a scientist".
Because it can process data across disciplines, the application scope of ScienceOne goes far beyond catalysts. In the field of mechanical engineering, ScienceOne has collaborated with the Institute of Mechanics of the Chinese Academy of Sciences to reduce the simulation analysis time of high-speed rail aerodynamic problems from several hours to a second-level response, and the error of key parameters in data-scarce scenarios has been reduced by 42%. In the field of astronomy, ScienceOne has collaborated with the National Astronomical Observatories to construct an intelligent inversion toolchain for stellar parameters, and the recognition accuracy of rare celestial bodies has increased by about 50% compared with existing methods. In the literature research process, which is the most time-consuming part of daily scientific research, a systematic literature review in the past usually took several days or even weeks, but ScienceOne 2.0 has reduced this process to about 20 minutes.
At present, ScienceOne has been deployed in more than 50 research institutes within the Chinese Academy of Sciences and more than 30 institutions outside the CAS. Its computing power system is based on the solution of Huawei Intelligent Computing Laboratory, and it has been adapted to chips such as Ascend.
These disciplines covered by ScienceOne have one thing in common: although their data are in different formats, they all have mature literature systems and formulas for reference, so AI modeling can be based on solid grounds.
There are still some disciplines that do not have such conditions. Their experimental data come from different types of equipment, with different formats and dimensions, and even a unified labeling standard has not yet been established, making it difficult for AI modeling to find a reliable starting point.
However, even under such conditions, AI has begun to produce serious scientific research results. For example, in the field of neuroscience.
In June 2026, a joint team of the Beijing Academy of Artificial Intelligence and Tsinghua University published a study in the journal "Science", which for the first time confirmed that memories can bidirectionally regulate sleep: positive memories make sleep more continuous, while negative memories exacerbate sleep fragmentation, providing new intervention ideas for sleep disorders associated with depression and anxiety.
The data analysis supporting this study is supported by the self-developed multimodal neuroscience large model of BAAI - Brainμ (pronounced Brain-mü).
Lei Bo, a researcher at the Beijing Academy of Artificial Intelligence, said at the roundtable forum that the reason why protein structure prediction is suitable for AI modeling is that the data standards of the entire field were already highly mature at that time, but neuroscience still does not have such a foundation today. Electroencephalography (EEG) records electrical signals on the surface of the brain, calcium imaging captures the activity of local individual neurons, and neural probes collect discharges from deep brain regions in limited areas. These signals come from different devices, different species, and different experimental paradigms, with different formats and dimensions. Lei Bo said that even for the same set of data, different laboratories may give completely opposite interpretations.
Facing such a data dilemma, Brainμ, self-developed by BAAI, has taken a different path, using a pre-trained multimodal model as a reference system. Lei Bo explained that the brain is a high-dimensional existence, and each detection device can only capture a low-dimensional projection of it, just like taking photos of the same building from different angles, each photo only contains part of the information. Brainμ uses the existing representation space of the multimodal models of language and vision to align the sparse data of each modality in neuroscience, allowing the model to learn a more complete brain representation than that of a single modality.
After the model masters this high-dimensional representation, it can realize the conversion between different modalities. For example, by inputting a segment of low-cost EEG data, it can output functional magnetic resonance imaging (fMRI) images with more information but requiring expensive equipment. For primary medical institutions lacking high-end equipment, doctors may be able to use AI to make diagnoses that could originally only be made by large tertiary hospitals.
The supporting BrainToken data platform integrates data from three species: humans, monkeys, and mice, covering major neuroscience modalities from EEG to two-photon calcium imaging. In cooperation with the National Institute of Biological Sciences, Beijing, Brainμ has completed automated labeling and analysis of more than 5,000 nights of mouse sleep data. The analysis results over 12 months have undergone two-way verification and are highly consistent with the manual judgments of professional sleep neuroscience doctors. Model training covers more than 70,000 nights of sleep records. According to the Beijing Academy of Artificial Intelligence, the team has collaborated with Huawei to complete the reasoning adaptation and optimization of Brainμ using Ascend supernodes, which fully supports the 12-month automated analysis process.
Lei Bo used the development stages of large language models to measure the current position of AI for Science. In his view, the model capability is roughly at the GPT-1 stage, the industrial sector's attention has reached the GPT-3 level, and scientists' usage feedback is close to the early stage of ChatGPT. For large language models, capabilities mature first, followed by applications. The logic of AI for Science is the opposite: applications are running ahead, pulling the models forward. "Demand is ahead of capabilities, and application validation in turn drives model iteration."
AI Starts Managing Laboratories
Conducting materials research requires different instruments to obtain information about the composition, structure, and morphology of materials. Ouyang Wanli, Vice President of Shenzhen Hetao College, said at the roundtable forum that just like a physical examination for people requires various devices such as electrocardiographs, B-ultrasound, and CT, the "physical examination" of materials is also inseparable from a series of precision instruments such as scanning electron microscopes, X-ray diffractometers, and energy spectrometers. The operation of these instruments is extremely dependent on humans. Researchers need to manually prepare samples, manually control the instrument software, repeatedly adjust parameters to wait for images to be generated, and then manually analyze the data. The software systems between different devices are mutually closed.
The AuraID multi-agent unmanned characterization platform launched by Shenzhen Hetao College in collaboration with Suzhou National Laboratory has chosen a different path. This system does not require instrument manufacturers to open software interfaces, but allows the agents to operate the device interface in the same way as human experts: observe the screen images, click buttons, read data, and adjust parameters.
In the test of the scanning electron microscope, after AuraID took over the operation, the noise in the initially captured images was obvious. The system automatically adjusted the acceleration voltage and magnification, the images gradually became clear, and finally stopped at the optimal parameters, automatically completing image capture and particle size analysis.
At present, AuraID has covered 6 types of precision instruments and more than 10 characterization modalities. In the micro-CT (micro-computed tomography) alignment task, the independent completion rate of AI has increased from 33% in version 1.0 to 80% in version 2.0, the workload of crystal structure analysis has been reduced by 50.6%, and the time spent on microstructure morphology analysis has been shortened from 9 minutes to 7.5 minutes. The same set of operation rules can support the characterization of three different types of samples: energy materials, nanomaterials, and metal materials.
Wang Zhihui, a representative from Shenzhen Hetao College, stated that automation on the synthesis side has been promoted by many scientific research institutions, but unmanned and intelligent operations on the characterization side are still a blank. AuraID is precisely entering this segment.
Yu Li, a professor at the School of Life Sciences of Tsinghua University and Director of the National Key Laboratory of Membrane Biology, introduced the implementation considerations of the "lights-out laboratory" in the field of life sciences at the roundtable forum.
In December 2025, Tsinghua University officially established the Artificial Intelligence Research Center for Life Sciences, providing an institutional foundation for the implementation of this direction. The concept of the "lights-out laboratory" originated from the "lights-out factory" in the manufacturing industry, where the production line does not require on-duty personnel and does not even need to turn on the lights. Yu Li introduced this idea into scientific research, allowing large models to be directly integrated into scientific research instruments, completing analysis while data is collected, and the results are entered into the database to feed back model training, forming a cycle that does not require continuous human intervention.
Yu Li revealed that his team began to introduce large models to assist the experimental process two years ago. The specific approach was to let students fully discuss the experimental plan with the large model before designing the experiment, to troubleshoot possible problems in advance. Last year, the team further deployed agents to manage the entire chain.
After introducing the large model, the experimental cost was reduced by 25% even when the team expanded by 20 people. Focusing on this direction, Yu Li has built supporting research institutions in Hangzhou and other places, equipped with infrastructure such as cryo-electron microscopes, large-scale computing power, and storage, bringing together chemists, biologists, doctors, and engineers to collaborate in the same space. According to Yu Li's vision, in the past, developing a new drug required 10 years and 1 billion US dollars. With the deep involvement of AI in the scientific research process, it may be reduced to 3 months and 3 million US dollars in the future.
The successful implementation of these achievements - ScienceOne's interdisciplinary modeling, Brainμ's brain science signal analysis, AuraID's instrument control, and Yu Li's lights-out laboratory - is inseparable from the support of computing power.
In addition, for domestic scientific research institutions, the importance of independent innovation in computing power should be further emphasized.
Scientific computing involves a large amount of sensitive data and core research results. Relying on a self-developed computing power base is not only a practical need to ensure data security, but also a basic condition for enhancing the overall competitiveness of domestic scientific research. The two solutions - Intelligent Computing Laboratory and Smart Scientific Research, launched by Huawei's Scientific Research Corps - have made explorations in this direction. The former solves computing power scheduling and data management, while the latter connects the entire scientific research process, and has now served more than 20 key laboratories across the country.
In the Tan Kah Kee Innovation Laboratory, the intelligent computing power platform for material science jointly built by Huawei has been put into operation, increasing the computational efficiency of tasks such as material simulation and molecular screening by 20%. From the supply of high-quality computing power to the digitization of the entire scientific research process, the self-developed software and hardware technology ecosystem is providing more and more complete support for the implementation of AI for Science.
AlphaFold has proved that in fields with the best data conditions, AI can achieve Nobel Prize-level results. What these four institutions are doing may not be that dazzling, but it is even more important. They want to enable AI to help in places where data conditions are far from ideal, so that scientists can spend less energy on repetitive work and leave more time for those problems that only humans can think through.
This article is from the WeChat official account "Economic Observer", written by Feng Kehan, and published with authorization from 36Kr.