In recent years, "Schools of Artificial Intelligence" have rapidly proliferated across Chinese universities. According to incomplete statistics, more than 50 universities have established such schools. Amid this wave, the Peking University Shenzhen Graduate School opted to create a school dedicated to "AI for Science," focusing artificial intelligence efforts on fundamental scientific challenges.

Traditional AI education has largely centered on cultivating talent in algorithms and engineering, while the goal of AI for Science is to use AI to reshape and accelerate the discovery process in basic sciences. Much like the impact AlphaFold has had on biology, this model seeks to transform AI from an application tool into a key driver of groundbreaking innovation in fundamental sciences such as physics, chemistry, and materials science.

"We believe that scientists engaged in fundamental research in mathematics, physics, chemistry, biology, materials science, and other fields should actively embrace AI and adapt to this shift in research paradigms—otherwise, their work is likely to fall behind." This was expressed by Tian Yonghong, Vice Dean of the Peking University Shenzhen Graduate School and Executive Dean of the School of Artificial Intelligence for Science, in a discussion with Intellectual, explaining the school’s rationale for choosing AI for Science.

The Intellectual: We’ve noticed that many universities are establishing artificial intelligence schools. Why did the Shenzhen Graduate School of Peking University choose the direction of "AI for Science"?

Tian Yonghong: There are two main reasons behind our decision to focus on AI for Science.

First, we observe a clear trend: artificial intelligence is profoundly transforming the paradigm of scientific research. A very typical example is protein folding prediction. Once AlphaFold emerged, it predicted almost all protein structures worldwide within just a year or two. This demonstrates the charm of AI. As a powerful tool, it is reshaping the mode of scientific research.

We believe that scientists engaged in fundamental research in mathematics, physics, chemistry, biology, materials science, and other such fields should actively embrace AI and adapt to this shift in research paradigms—otherwise, their work is likely to fall behind.

Second, this decision aligns with the overall development strategy of Peking University. Over the past two decades since its establishment, the Peking University Shenzhen Graduate School has cultivated an academic environment rich in interdisciplinary collaboration, providing an innovative testing ground for exploring new models of AI for Science. With Peking University’s main campus boasting strong foundational disciplines such as mathematics, physics, and chemistry, we aim to pilot these explorations in Shenzhen. If successful, these valuable experiences and models can later be extended to the main campus and even across the country, serving as a pioneering effort.

The Intellectual: In your opinion, what is the biggest bottleneck currently restricting the rapid development of "AI + Science"? To what extent can the establishment of the School of Scientific Intelligence change this situation?

Tian Yonghong: The core issue to address is the talent gap.

In many traditional industries, like pharmaceuticals, everyone recognizes the importance of AI, but few know how to actually implement it. For example, during our research visits to some companies, we found that everyone from the directors to the researchers was aware of the need to use AI tools. But in practice, many were still relying on traditional statistical learning methods from a decade ago. When it comes to new technologies like large language models, they understand their significance, yet remain unsure about how to apply them.

Even leading companies face such dilemmas, which shows how sorely lacking enterprises across the board are in interdisciplinary talents who understand both AI and basic science. We hope that by establishing this college, we can cultivate a large number of such talents to address the existing shortcomings.

The Intellectual: In the past, many "AI for Science" projects were often carried out through interdisciplinary collaborations within specific laboratories or research teams. Why is it necessary now to systematize AI for Science and establish a dedicated college to cultivate interdisciplinary talents? How does this training model differ from traditional interdisciplinary collaboration?

Tian Yonghong: To be frank, we have been following the traditional interdisciplinary collaboration model for many years, but there have always been some persistent challenges that are difficult to resolve.

First is the issue of collaboration. For instance, if a biology professor wants to work with an AI professor, the latter might already be overwhelmed with their own projects and lack the bandwidth to participate—such situations are very common. Even if they do collaborate, the second challenge is determining the ownership of intellectual property. After the project is completed, who becomes the first author of the paper, and who becomes the corresponding author? To this day, this remains a difficult problem to resolve in academia.

Scientists often believe that since they define the research questions and provide the data, they should take the lead. Meanwhile, AI researchers argue that they contribute the essential tools and methodologies, and thus deserve due recognition. This conflict is difficult to resolve.

Most crucially, for students engaged in such collaborations, which academic committee should evaluate and approve their interdisciplinary theses? The traditional disciplinary system is fragmented—mathematics, chemistry, and computer science each have their own committees and evaluation criteria. An interdisciplinary thesis may not be fully understood by reviewers from either side, and it can be challenging to find appropriate evaluators. As a result, students may face difficulties in graduation and obtaining fair assessments.

The Intellectual: How does the model of the School of Artificial Intelligence differ from traditional interdisciplinary collaborations?

Tian Yonghong: We implement a dual-supervisor system. Two professors from different disciplines share joint responsibility from the very beginning, and their shared achievement is cultivating talent in AI for Science. This naturally resolves the issue of willingness to collaborate. With two supervisors guiding one student, intellectual property matters can also be more easily addressed through negotiation. The student’s affiliation is clear, eliminating disputes over first authorship or corresponding authorship.

The college will establish a dedicated interdisciplinary degree evaluation committee to review such cross-disciplinary theses. This represents an institutional shift from the previous model that relied on individual self-study or fragmented collaborations, enabling the systematic and institutionalized cultivation of interdisciplinary talent, thereby genuinely advancing the development of AI for Science.

The Intellectual: You just mentioned the dual-supervisor system. How is the division of responsibilities structured in its actual implementation?

Tian Yonghong: The dual-supervisor system is a core part of our approach tailored to current realities. Frankly, there are very few scientists worldwide who are well-versed in both AI and Science. It’s difficult to find enough single supervisors capable of mentoring such students independently. Therefore, we adopt a model where two supervisors jointly guide one student. Ideally, these two supervisors should have a prior collaborative foundation, or at the very least, a strong willingness to work together.

During the mentoring process, their division of labor is clear. The scientific supervisor is responsible for setting the direction—identifying which scientific questions are worth pursuing and which represent long-standing challenges in the field that have remained unsolved for decades. The AI supervisor, on the other hand, is tasked with determining the technical approach, analyzing whether the scientific problem can be accelerated or resolved using AI technology.

By combining their expertise to co-supervise students, it also means that both supervisors must be present during student progress reports. This system increases the time commitment required from the mentors, while students benefit from more learning hours and a richer curriculum. Such a unique training and research experience ultimately fosters stronger competitiveness.

The Intellectual: When it comes to evaluation, there has been a perception that interdisciplinary research tends to be at a disadvantage within the existing academic evaluation system. As a newly established school, how will the School of Artificial Intelligence for Science reform the research evaluation system to avoid being assimilated by the current framework?

Tian Yonghong: Research evaluation acts as a baton—if the baton is not wielded properly, interdisciplinary studies will inevitably suffer. While it’s beyond our school’s capacity to completely resolve this long-standing issue on our own, we are currently presented with a valuable opportunity for reform. There is now a broad societal recognition that advancing interdisciplinary research is essential, and the academic community as a whole is becoming more inclusive toward such initiatives.

Within our school, we have already undertaken some initial explorations. First, regarding student evaluation, Peking University has specifically established an interdisciplinary degree assessment committee, which reviews the theses of our AI for Science students. This committee evaluates degree papers from an interdisciplinary perspective.

Second, in terms of recognizing faculty research achievements, the school takes the lead in its academic field by acknowledging equal contributions from co-first authors and co-corresponding authors, giving them equal consideration in workload calculation and achievement recognition. We believe that in any AI for Science research, contributions from both the scientific and AI aspects are equally important. Such outcomes can only be achieved quickly and effectively through close collaboration between the two sides, and it is essential to abandon the traditional notion of primary and secondary roles.

Of course, our school’s exploration is only partial; broader reforms to the evaluation system still depend on the state adjusting its "baton." This will be a long-term process—it cannot be achieved overnight. But through such explorations, we hope to gradually guide the formation of a new research culture that recognizes interdisciplinary integration.

The Intellectual: In terms of the curriculum design, which disciplines will AI intersect with?

Tian Yonghong: In curriculum development, our initial interdisciplinary focuses are primarily on the integration of AI with physics, chemistry, life sciences, and materials science. Of course, this is just our starting point. In the future, the application of AI will permeate more STEM disciplines—such as energy and environment—and even extend into the humanities and social sciences.

In our current educational framework, the curriculum is structured around "one core with six directions." The "one core" refers to a common foundational course platform required for all students. This platform includes three types of courses: first, introductory courses such as Introduction to Artificial Intelligence and Introduction to Artificial Intelligence for Science; second, computing and intelligence courses covering machine learning, scientific computing, and programming; and third, interdisciplinary foundation courses like First Principles of Science and scientific research methods.

After completing this "core curriculum," students then choose from different specialized elective modules based on their interests and academic direction—what we refer to as the "six directions." For example, those who choose the AI for Life Science track will take courses such as computational biology, proteomics, and AI-based drug design, while students specializing in AI for Materials will study subjects like computational materials science and material simulation.

The Intellectual: In terms of concrete platform development, how does the School plan to implement "AI for Science"?

Tian Yonghong: Research and education in AI for Science differ significantly from traditional AI. In traditional AI training, a student might only need computing platform support. However, we require not only computing power and data but, more crucially, an integrated "dry-wet lab" experimental platform. Here, "dry lab" refers to the design and computation of AI models, while "wet lab" involves real-world physics, chemistry, and biology experiments. We aim to closely integrate these two aspects, enabling students to train on such a platform during their studies and master this new paradigm of scientific research.

For instance, to design a new drug molecule, we first use large AI models on computers to generate candidate molecular sequences, screen the most promising ones, and automatically generate experimental protocols. Automated robotics then carry out the synthesis and validation. These automated setups not only boost efficiency but also lower the barrier for AI students to handle complex experiments. Once the experiment is completed, data is automatically sent back, visually analyzed, and fed into the model, forming a highly efficient closed loop for automated iteration. Such facilities and environments are absent in traditional AI or conventional laboratory settings.

The Intellectual: In terms of data resources and computing power sharing, how will the School promote the infrastructure development for "AI for Science"?

Tian Yonghong: In building such a platform, our goal is not only to provide a research infrastructure for AI for Science but, more importantly, to promote the sharing of scientific data. For a long time, many scientists have been accustomed to keeping their data tightly controlled, resulting in the formation of thousands of fragmented and isolated "data silos."

However, in future big science research, true breakthroughs can only be achieved by integrating scattered data. Teams that take the lead in collecting, organizing, and sharing big science data will have a greater chance of producing significant outcomes. In contrast, adhering to a "data privatization" model will considerably slow down research progress and may even hinder substantive achievements. This trend will also drive a fundamental shift in the data-sharing paradigm within the scientific community.

I see this as a major opportunity. Building such new infrastructure is foundational work with long-term value—much like constructing high-speed railways or developing electric vehicles—and is of great importance to the nation’s scientific and technological advancement.

The Intellectual: The School is located in the Guangdong-Hong Kong-Macao Greater Bay Area, which boasts a highly dynamic innovation culture and a cluster of enterprises. How will the School collaborate with the industry?

Tian Yonghong: We are currently exploring collaborations with several companies in various forms. For instance, we may jointly establish experimental platforms or adopt an "industry-proposed problem" approach, where enterprises provide real industrial challenges as practical training topics for our students. Our role is to refine and elevate these practical problems brought by companies, stripping away specific engineering details to identify the most fundamental scientific or technological challenges, which our faculty and students will then tackle.

The Intellectual: Many people are now discussing whether AI could potentially autonomously discover new scientific laws, rather than merely performing efficient predictions and calculations within existing frameworks. What are your thoughts on this?

Tian Yonghong: This can be considered one of the ultimate goals of AI for Science research. With technological advancements, this possibility will undoubtedly become a reality. At the current stage, however, the primary role of AI remains accelerating scientific discovery.

It is reflected in the "dry lab" phase, where we use large models to efficiently explore synthesis pathways, and also in the "wet lab" phase, where automated robots conduct experiments in parallel with far greater efficiency than manual work. In the past, producing one set of data on carbon nanotube materials might have taken us a full day, but now, robots can generate ten sets overnight with highly reproducible data. This speed alone significantly accelerates scientific research.

In fact, the seeds of AI’s autonomous discovery have already begun to emerge. Our next step is to enable AI models, by learning from vast amounts of scientific data, to further develop the ability to propose scientific hypotheses and even conduct autonomous exploration of certain scientific questions. I believe this day will come sooner rather than later. Of course, solving deeper and more challenging problems, such as Goldbach’s Conjecture, will still require a very long time.