2023 Zhongguancun Forum | When AI meets scientific research, scientists have something to say

On August 10, 2018, Peking University organized an in-school discussion and proposed the concept of AI for Science (scientific intelligence) for the first time. The Beijing Institute of Scientific Intelligence was established last year and is the world's first research institution with the theme of "AI for Science". In fact, the academic community has carried out relevant scientific research earlier and has accumulated certain results. On May 30, at the 2023 Zhongguancun Forum "Artificial Intelligence Driven Scientific Research Forum", a number of domestic and foreign experts conducted in-depth discussions and exchanges on topics such as the importance of AI for Science, research results and talent training, and focused on the role of artificial intelligence in science and technology. Research progress and breakthroughs in life sciences, materials science and other fields will be shared in order to further tap the potential of artificial intelligence in scientific research.

Why is it important

Let the scientific research system transition from workshop mode to "Android" mode

In recent years, AI for Science has formed a consensus among leading international research institutions, and countries are increasingly paying attention to the huge potential of AI for Science. The reason why this concept is popular among scientific research institutions starts with the bottleneck of scientific research.

"There are two ways to do scientific research, data-driven and basic principle-driven." E Weinan, an academician of the Chinese Academy of Sciences and director of the Beijing Institute of Scientific Intelligence, introduced the difficulties in these two models in detail, "The bottleneck of data-driven is data The collection efficiency is low and there is a lack of effective data analysis methods. As for driving by basic principles, Paul Dirac, one of the founders of quantum mechanics, said a hundred years ago that the task of seeking basic principles has been basically completed, but using basic principles to solve practical problems The problem is less efficient because the mathematical problems expressing the underlying principles are too difficult.”

This results in simple problems such as structural mechanics, mechanical engineering, aerospace, and electronic engineering that can be solved. Complex problems such as material properties and material design, drugs, and catalysts can only be solved theoretically through experience and trial and error. and application separation.

What is the boundary between simple and complex problems? E Weinan believes, "It is the number of degrees of freedom (dimension), which means that as the number of dimensions increases, the complexity index of the problem increases. This is what AI can help us solve. AI provides data-driven models New tools can improve the reliability and efficiency of fundamental-principle-driven models, and can also combine data-driven and fundamental-principle-driven models."

The four basic tools of scientific research are basic principles and data analysis methods, experiments, literature, and computing power.

From the perspective of new scientific research infrastructure construction, "The specific operation method for us to do scientific research is still the small farmer workshop method. For example, to conduct experiments, buy all the equipment and do it from beginning to end. This cycle is very long and inefficient. ", E Weinan said that AI will promote the construction of the next generation of tools, transition from the past workshop model to the Android model, and build a large scientific research platform, an open source platform for basic principles, a knowledge base for literature, and experiments. Several experimental centers and an experimental cloud platform, several computing centers and a computing cloud platform for computing power.

What are the results

Literature knowledge base based on large language model vector database

The consensus is there, and the tools are highly anticipated. According to statistics from the National Science Foundation, researchers spend 51% of all scientific research time on searching and digesting scientific and technological information, 8% on planning and thinking, 32% on experimental research, and 9% on written summaries. The scientific research retrieval method will move from the consultative eye-checking and hand-turning and Internet-based search retrieval stages to the conversational retrieval stage.

“But retrieval through conversational large language models has limitations, such as illusions and biases, data lag, and cache limitations,” Meng Zhuofei, vice president of Moqi Technology, concluded.

Based on these pain points, the Beijing Institute of Scientific Intelligence, the Computer Network Information Center of the Chinese Academy of Sciences, and Moqi Technology jointly released Science Navigator V1.0, a literature knowledge base based on large models + vector search engines. The knowledge base includes The user layer represented by scientific researchers, the model layer represented by GPT4, etc., the middle layer composed of vector database search engines, and the data layer supported by literature and teaching materials. The model layer is in charge of the Beijing Institute of Scientific Intelligence, the middle layer is in charge of Moqi Technology, and the data layer is in charge of the Computer Network Information Center of the Chinese Academy of Sciences.

Meng Zhuofei introduced that ScienceNavigator V1.0 supports cross-modal recognition processing of text, pictures, tables, and formulas. Scientific researchers can use various large and small models such as Wen Xinyiyan and LLM to achieve the best results in problem analysis. First The batch has included nearly one million papers in the fields of chemistry, materials, AI and other fields, and will be expanded to hundreds of millions of documents in the natural and humanities disciplines in the future.

How to vertically integrate:

Establish a collaboration system + promote AI for Science into the classroom

AI for Science means interdisciplinary and large-scale integration. Its rapid development will trigger the reconstruction of scientific research models. There are still many challenges and problems in terms of talents, mechanisms, ecology, interdisciplinary and other aspects. How to build vertically integrated scientific research in the new era? The system requires collective effort.

Dr. Huang Tiejun is the president of Beijing Zhiyuan Artificial Intelligence Research Institute. He has more than 30 years of scientific research experience. His most profound understanding is collaborative cooperation. "Regardless of application, research and development, or basic theory, how to form a collaborative system is particularly important. Regarding the application of AI in the scientific field, we need to think about how to cooperate better and achieve results with higher efficiency. For example, there are more than 60 kinds of elementary particles, There are more than 100 kinds of atoms, and the commonly used numbers are relatively clear. If the organization is good, you can go through the entire periodic table of elements, and gradually build up the entire basic model system of physics, chemistry, and life, and everyone can do research and development and applications on it. The efficiency will be much higher.”

Lei Lei, assistant to the dean of the School of Materials Science and Engineering at Peking University, who is responsible for teaching undergraduate and graduate students, emphasized talent cultivation, “This year our college launched a series of AI for material science courses, inviting our school and other colleges and universities of Peking University and Relevant experts come to class. We hope to promote the concept of AI for Science into the classroom as soon as possible, so that students can break away from the original way of thinking and tools and use new tools as soon as possible. New tools really help accelerate the progress of experimental science, and the enthusiasm of students and teachers The evaluations are all very good. Some students and teachers put forward new suggestions, for example, the relevant content is still not rich enough and more technical reference materials are needed.

▲Micro Classroom

AI for Science Material Science

"Materials are the driving force behind almost all technologies," said David Srolovitz, a member of the National Academy of Engineering in the U.S. and the Dean of the Faculty of Engineering at the University of Hong Kong. Artificial intelligence-based interatomic potential energy enables large-scale atomic simulations with near-QM accuracy. The specialized application of DP (Data Processing) reveals the perfect crystalline and defective properties of structural metals and alloys, so DP is the only way to understand the properties of many defects. ”

AI for Sciencelife science

Xu Jinbo, Distinguished Visiting Professor of Tsinghua University and Professor of Toyota Computing Technology Institute in Chicago, USA: "Protein is the material basis of life and the main carrier of life activities. AI technology is suitable for de novo protein design in various application scenarios, such as designing to reduce The toxic new protein drug can be used to treat various types of tumors such as gastric cancer, colorectal cancer, liver cancer, melanoma, etc. Green pesticides can be designed to ensure food safety and security."

AI for ScienceAtomic Dynamics

Robert Carr, academician of the National Academy of Sciences and professor of the Department of Chemistry at Princeton University: “Without empirical input, AI modeling of complex molecular processes from basic quantum theory can have good/excellent predictive capabilities, and can be used in difficult-to-understand situations. This approach is important in chemistry, materials science, and biology when chemical reactions are described by empirical models. With more precise reference quantum mechanical models, systematically improving accuracy should be possible."

Beijing Business Daily reporter Wei Wei

The above is the detailed content of 2023 Zhongguancun Forum | When AI meets scientific research, scientists have something to say. For more information, please follow other related articles on the PHP Chinese website!