What Is AI for Science? How AI Is Changing Scientific Research, and Japan's National Strategy, Explained Clearly

This is Hamamoto from TIMEWELL.

AI for Science is a term for a new way of doing research: building artificial intelligence into science not only as a tool for running the work faster, but as something that poses hypotheses, designs experiments, and produces discovery itself. AI now names the shape of a protein in minutes, and that work became the 2024 Nobel Prize in Chemistry. At the International Mathematical Olympiad, AI earned gold-medal-level scores, and in Japan, RIKEN has launched a supercomputer dedicated to science called "Rikyu," while the government has begun investing on the order of hundreds of billions of yen. In this article, I lay out what AI for Science means, what is happening around the world, how Japan is moving as a nation, and how high school inquiry-based learning can serve as an entry point — using only primary sources from the developers and the government.

Our core business at TIMEWELL is hands-on AI support for companies, but I have had more and more chances to hear from high school teachers asking how to handle the story of frontier science and AI in an inquiry class. It is a fascinating topic, yet a well-organized set of primary sources is surprisingly hard to find. Let me start with three key points.

• AI for Science is not just about "using AI to make research more efficient." It points to a shift toward a research paradigm in which AI becomes a driver of discovery. That protein structure prediction became a Nobel Prize in Chemistry is the symbol of this.
• Around the world, real examples of AI moving the genuine frontier came in quick succession from 2024 to 2025: AlphaFold 3, the genome foundation model Evo 2, the algorithm-discovery agent AlphaEvolve, and gold-medal-level results at the math olympiad.
• Japan is following suit at a national scale, with RIKEN's "Rikyu" supercomputer and MEXT's promotion measures (on the order of 114.3 billion yen in the FY2025 supplementary budget). High school inquiry-based learning is the first step by which students can connect to this larger current.

What is AI for Science, and why "now"?

Using AI in science is not, in fact, new. Sifting through vast observational data or speeding up simulations has long been part of research. What AI for Science refers to is a step beyond that. Generating a hypothesis, deciding which experiment to run next, designing a molecule or a sequence that no one has ever seen. AI has begun to enter the part of research that human scientists have always owned — the judgment that moves toward discovery. That paradigm itself is what we call AI for Science.

In an official document, MEXT writes that building AI into scientific research brings "dramatic improvements to the scope and speed of research" and "a rapid and fundamental transformation to the very nature of scientific research"¹. That is fairly bold language for a government paper. The same document names AlphaFold, which predicts protein structures, as a concrete example, praises it for "dramatically reducing the time and cost that research takes," and positions AI as not merely raising research productivity but "transforming the very nature of scientific research"¹. The state, naming one particular AI, is acknowledging that the tide of science has turned. That is worth noting.

So why "now"? I see two main reasons. One is that, with the arrival of generative AI and foundation models, AI can now reason in natural language and handle knowledge across specialized fields. The other is that AI's "persistence" has grown. Until recently, AI was something that answered a question in seconds. Now, agent-style AI has emerged that spends tens of minutes to several hours — sometimes days — investigating and thinking a problem through. AI that settles in to explore can do more than solve problems that already have answers; it can reach into research territory where the answer does not yet exist. I go deeper into how this long-running AI affects research in the article on AI-driven research.

The global frontier: AI is already moving the leading edge of science

Abstractions alone do not sink in. Here are the cases, with dates and developers, where AI actually produced results at the frontier of science between 2024 and 2025. I have limited these to the developers' official announcements or facts confirmed by public institutions.

Announcement	When	Who	What happened
AlphaFold 3	May 8, 2024	Google DeepMind / Isomorphic Labs	Predicts the structure and interactions of proteins plus DNA, RNA, and drug molecules2
Nobel Prize in Chemistry	October 9, 2024	David Baker, Demis Hassabis, John Jumper	Awarded for protein design and structure prediction3
Evo 2	February 19, 2025	Arc Institute / NVIDIA	Released a genome foundation model that learned the DNA of life at scale4
AlphaEvolve	May 14, 2025	Google DeepMind	Updated best-known solutions on unsolved and difficult math problems5
Gold-medal-level math olympiad	July 2025	Google DeepMind / OpenAI	Achieved gold-medal-standard scores at the International Mathematical Olympiad6

Life science: AI reads proteins and genomes

Life science became the symbolic stage for AI for Science. AlphaFold 3, announced on May 8, 2024, predicts not only the three-dimensional structure of proteins but how they combine with DNA, RNA, and the small molecules that become drug candidates². The developer, Google DeepMind, states that for predicting interactions between molecules it delivers "at least a 50% improvement over existing methods" (this figure is the developer's own official claim)². Determining the shape of a single protein by experiment once took years. That work is being replaced by predictions that take minutes.

What put this current on the world stage was the Nobel Prize in Chemistry on October 9, 2024³. Half the prize went to research on designing proteins computationally (David Baker), and the other half to protein structure prediction by AlphaFold (Demis Hassabis and John Jumper). AI-driven science became that year's chemistry prize itself. I think it is the clearest possible evidence that AI is doing star-level work at the frontier of research.

One more thing stirs the imagination: genome foundation models. Evo 2, released on February 19, 2025 by Arc Institute, NVIDIA, and others, learned the genomes of a wide range of organisms — from bacteria to eukaryotes — as if they were "text," at massive scale⁴. The volume of bases it learned exceeds 9.3 trillion. As a result, the developers report that it can distinguish whether a variant of the breast-cancer-related gene BRCA1 is "benign or pathogenic" with over 90% accuracy⁴. AI that has become able to read DNA, the language of life, is reaching into disease-risk assessment and the design of new sequences.

Math and algorithms: updating records humans could not break

It is not only life science. Even in mathematics, long held up as a symbol of human intellect, AI has rewritten actual records. AlphaEvolve, announced by Google DeepMind on May 14, 2025, took on more than 50 unsolved and difficult math problems, rediscovering the best-known solution in about 75% of cases and updating the previous best in 20%⁵. A symbolic example is a geometry problem, the "kissing number in 11 dimensions," studied for over 300 years, where it raised the lower bound from 592 to 593⁵. It even improved the procedure for multiplying 4×4 matrices, which had been considered the best since 1969⁵. AI updating records unbroken for decades, using solution methods it found on its own. This one genuinely surprised me.

The math olympiad story is also not to be missed. In July 2025, Google DeepMind announced that its model achieved 35 out of 42 points at the 2025 International Mathematical Olympiad, fully solving 5 of the 6 problems and meeting the gold-medal standard⁶. What is more, this result was officially graded and certified by the IMO organizers⁶. It was groundbreaking that the model wrote proofs in natural language rather than a specialized formal language, and solved within the same time limits as humans. Here, one caveat is needed. Around the same time, OpenAI also announced gold-medal-level results on the same problems, but this was the company's own evaluation rather than an official grading by the organizers⁶. It is easy to lump these together as "AI won a gold medal," but the distinction that only DeepMind's result was officially certified is worth keeping straight if you plan to cover it in class.

Lining up results like these makes one want to exaggerate, but expressions such as "world's first" or "largest" from the developers are quoted strictly as the developers' claims. They are not necessarily facts that third parties have fully verified. Especially when handling science, I try not to let that line blur.

Japan's national strategy: RIKEN's "Rikyu" and MEXT's 114.3 billion yen

The global story is almost overwhelming, but Japan is not standing idle. If anything, it is moving quite seriously, at a national scale.

The "Rikyu" supercomputer: a computing base dedicated to science

The symbol of this is "Rikyu," a supercomputer for scientific research launched by RIKEN. The name was decided on June 19, 2026⁷. Its development and operation are handled by the Scientific Foundation Model Development Program (known as AGIS) within RIKEN's Transformative Research Innovation Platform (TRIP) headquarters, together with the Center for Computational Science (R-CCS)⁷. The system is built from more than 400 computing nodes carrying NVIDIA's latest chips, and it aims, in coordination with Japan's flagship supercomputer "Fugaku," to build a world-class development environment for AI for Science⁸. Operations are scheduled to begin in July 2026⁷.

The origin of the name is elegant. It is a coinage meaning to explore the "ri" (the underlying principles of natural phenomena) with AI and computing, and then to "master" (kyu) them — and it deliberately echoes the tea master Sen no Rikyū in sound⁷. As a name for Japanese-born science AI, I sense it was chosen with an eye to being easy to convey worldwide. It may look like the story of building a single machine, but it is really a statement of intent: to grow AI for science domestically. I have also prepared a detailed explainer on RIKEN's AGIS and the "Rikyu" supercomputer, digging into TRIP-AGIS, the partnership with Argonne National Laboratory, and the relationship to Fugaku NEXT, all from primary sources.

MEXT's promotion measures: a strategy and a budget in the hundreds of billions

The outlines of the institutions and the budget have also come into focus. MEXT set up an AI for Science Promotion Committee and, on March 31, 2026, formulated a "Basic Strategy for Promoting AI for Science"⁹. The strategy sets out a direction of growing scientific foundation models domestically, in addition to Japan's strong fields such as life science and materials. A line in the national document even reads that "maintaining domestic AI research and development capability is extremely important from a security standpoint"¹, betraying a sense of urgency about not leaving AI for science to other countries.

The scale of the budget is concrete. For the "innovation of scientific research through AI for Science," 114.3 billion yen was allocated in the FY2025 supplementary budget, or 152.7 billion yen including related costs¹. The FY2026 initial budget proposal also secures 19.3 billion yen¹. At the core is the 37.0 billion yen "Program for Scientific Research Innovation through AI for Science," broken down as follows.

Category	Budget scale	Content
Project type	32.0 billion yen	Concentrated investment in priority fields that are Japan's winning cards. About 2 billion yen per project, in principle over 3 years1
Challenge type	5.0 billion yen	Backing researchers across every field. About 5 million yen each for roughly 1,000 projects, generally within a year1

This challenge type is the open-call program known as "SPReAD." Covering every field, including the humanities and social sciences, it is set to distribute research funding of up to 5 million yen per project to around 1,000 projects, with the first call held from April 17 to May 18, 2026¹⁰. The flow works like this: the state establishes a fund at the Japan Science and Technology Agency (JST), which then delivers funding and computing resources to universities and others¹. Spread small grants widely to cultivate the base, while investing big in priority fields. It is a two-track design aimed at lifting research capability from the bottom up.

This whole set of moves is positioned as driving the improvement in research capability sought by the 7th Science, Technology and Innovation Basic Plan, the overarching framework of national science and technology policy¹. The hardware of Rikyu and the software of strategy and budget have begun to point in the same direction. Honestly, that this much has taken shape in just a few years was faster than I expected.

How to reach the entry point of AI for Science through inquiry-based learning

So far this has been a story of researchers and the state. So how does it connect to the high school classroom? I believe inquiry-based learning is the closest, most accessible entry point to AI for Science.

The reason is that the two processes are strikingly alike. Japan's Courses of Study embed in the goals of the Period for Inquiry-Based Cross-Disciplinary Study a process of "finding a question from your own relationship with real society and real life, setting your own task, gathering information, organizing and analyzing, and summarizing and expressing." Pose a question, gather information, analyze, express. Structurally, this is the same as what the researchers behind AlphaFold and Evo 2 are doing. Only the scale and the tools differ; the skeleton of the work is the same. That is exactly why an inquiry class can become a place to experience "AI for Science in miniature." I write about the concrete ways to combine inquiry and generative AI, down to the steps, in the practical article on high school inquiry-based learning × generative AI.

What can you actually do? For example, have students open the protein structure database that AlphaFold publishes and look at the shape of a protein from a familiar organism. Have them use generative AI as a sparring partner, asking "where is this hypothesis weak?" to sharpen the question. Have a long-running research-style AI investigate across primary sources, and have the students verify the result. Every one of these was, a few years ago, a tool only researchers could touch. If you are unsure how to set a theme, a collection of AI-centered inquiry themes may help.

That said, get the order wrong and it all falls apart. Let AI take over the question itself, and inquiry turns into "having AI produce the answer and cleaning it up." MEXT's N-E.X.T. High School Vision, too, sets out a two-stage approach: not shutting AI out, but first cultivating abilities AI cannot replace, then mastering the use of AI. I have organized this thinking in detail in the guide to the N-E.X.T. High School Vision and in the article that digs into "abilities AI cannot replace". Students keep hold of the role of posing the question. AI stays a partner for verification and acceleration. That line is what turns inquiry into an entry point to AI for Science.

At TIMEWELL, too, through WARP for Schools, our program for schools and educational institutions, we help design inquiry-based learning that uses generative AI and provide hands-on support in which students build products with their own hands. We have been involved in developing more than 500 people and have worked on a partnership project with the Tokyo Metropolitan Government (WARP ENTRE). We want to deliver not just instruction in operating AI, but the experience itself of "posing your own question and giving it form." If you work at a school and want to think together about how to connect the frontier of science and AI to classroom inquiry, please take a look.

Summary

• AI for Science is a new research paradigm that builds AI into research not merely as a tool for efficiency, but as a driver of hypotheses, experiment design, and discovery.
• Around the world, real examples of AI moving the frontier came in quick succession from 2024 to 2025: AlphaFold 3, the 2024 Nobel Prize in Chemistry, Evo 2, AlphaEvolve, and gold-medal-level math olympiad results.
• Japan is following suit at a national scale, with RIKEN's "Rikyu" supercomputer (operations planned for July 2026) and MEXT's promotion measures (on the order of 114.3 billion yen in the FY2025 supplementary budget).
• High school inquiry-based learning shares the same skeleton — "pose a question and verify it" — as AI for Science, making it an entry point through which students connect to the frontier.
• The key is that students never let go of the role of posing the question. A design that avoids outsourcing everything to AI is what turns inquiry from mere busywork into genuine inquiry.

The frontier of science no longer belongs to researchers alone. High schoolers in an age when a protein's shape appears in minutes will learn while glimpsing discoveries out of the corner of their eye, before they reach the textbooks. Standing closest to that entry point, I believe, is the inquiry class. That first step can begin with today's choice of theme.

---

References

Related articles

• What Are RIKEN's AI Science Hub AGIS and the "Rikyu" Supercomputer? Japan's National Science-AI Strategy, Explained
• What Is AI-Driven Research? How Long-Running Agents Are Reshaping Research
• What Is the N-E.X.T. High School Vision? A Clear Guide to MEXT's High School Reform
• What Are "Abilities AI Cannot Replace"? How High School AI Education Will Change
• High School Inquiry-Based Learning × Generative AI: Practical Steps to Empower Students
• 50 AI Inquiry-Learning Themes: A Ready-to-Use Collection for Teachers

Footnotes

1. MEXT, "On the FY2025 Supplementary Budget and FY2026 Initial Budget Proposal for AI for Science" (AI for Science Promotion Committee, Document 2, February 9, 2026) ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹

2. Google, "Google DeepMind and Isomorphic Labs introduce AlphaFold 3" (May 8, 2024) ↩ ↩² ↩³

3. Google DeepMind, "Demis Hassabis and John Jumper awarded Nobel Prize in Chemistry" (October 2024) ↩ ↩²

4. Arc Institute, "Evo 2: A biological foundation model" (February 19, 2025) ↩ ↩² ↩³

5. Google DeepMind, "AlphaEvolve: a Gemini-powered coding agent for designing advanced algorithms" (May 14, 2025) ↩ ↩² ↩³ ↩⁴

6. Google DeepMind, "Advanced version of Gemini with Deep Think officially achieves gold-medal standard at the International Mathematical Olympiad" (July 2025) ↩ ↩² ↩³ ↩⁴

7. RIKEN, "Name of the AI for Science Development Supercomputer Decided as 'Rikyu'" (June 19, 2026) ↩ ↩² ↩³ ↩⁴

8. RIKEN, "System for the AI for Science Development Supercomputer Decided" (July 28, 2025) ↩

9. MEXT, "Basic Strategy for Promoting AI for Science" (March 31, 2026) ↩

10. MEXT, "Program to Create Emerging Challenge Research through AI for Science (SPReAD)" special site ↩