CONTENTS

Complexity: the limits of human science

My own background is in physics. Broadly speaking, physics has two major directions. One seeks to push deeper into fundamental laws, such as those represented by elementary particle physics. The other asks why such diverse and complex phenomena emerge in the world from those basic laws. I have long been interested in the latter question: how complex entities, including living systems, come into being.
But once complexity itself becomes the object of study, the traditional modes of theoretical understanding quickly reveal their limits. Problems that can be solved elegantly with paper and pencil make up only a small fraction of nature. Many nonlinear phenomena, shaped by intertwined interactions, cannot be solved analytically. One must rely either on simulation or on building models directly from data. My own work has long centered on simulation, but the situation has begun to change markedly in recent years.
The effort underway at AGIS to accelerate science with artificial intelligence, often described as AI for Science, is based on this shift. AI for Science is not only for improving research efficiency: faster calculations, automated experiments and the like. What is more important is that AI is beginning to accelerate the core processes of science itself, such as hypothesis generation and model building.

How data changes science

In scientific research so far, hypotheses have depended heavily on the experience and intuition of human experts. Experiments, theory, simulation, and observation have been made to gradually open up unknown territories of knowledge. That back-and-forth cycle has been one of the basic structures of science.
AI introduces an entirely different element into that structure. It can handle vast amounts of data across domains, extract high-dimensional patterns that humans cannot grasp all at once, and produce abstractions that cut across fields. These capabilities raise the possibility that AI could suggest hypotheses that do not simply extend human experience or intuition.
A striking example is AlphaFold, an AI system for protein structure prediction. AlphaFold made it possible to infer three-dimensional protein structures with high accuracy, a process that had long depended heavily on experimental methods. AlphaFold emerged not only because machine-learning methods improved, but also because protein structure data had accumulated over many years. Even if today’s machine-learning techniques had existed 20 years ago, the available data at the time would probably not have supported the same achievement.
In AI for Science, the way data is acquired and accumulated is therefore a decisive factor in determining whether research succeeds. That, in turn, makes the automation of experiments and observations, including robotic systems for data generation, indispensable. One of AGIS’s major goals is to build foundation models for science. But in parallel, we are also working across experimental science, life science, materials science, computation, and robotics to integrate AI and automation into the research process as a whole.

The scientist of the future will be a human-AI hybrid

One of the central challenges in AI for Science is to redefine the agent of science. Traditionally, that role has been occupied by the human scientist. AI was a tool, no more than an auxiliary presence. But as AI becomes more deeply involved in generating hypotheses and constructing models, the agent of science will shift from the human alone to a composite of human and AI.
This does not mean that scientists will become unnecessary. Decisions about what questions are worth asking, where value lies, how to interpret the outputs produced by AI, and how to position them in society will place even greater demands on human judgement.
The scientist’s task is likely to be reshaped with the advancement of AI. Scientists will spend less of their time manually operating apparatus and more of it designing research as a whole. Structuring the question, determining which parts AI should handle, and deciding where human judgement must intervene will become the central scientific work.

Science as an experimental site of transformation

Consider a thought experiment. An AI makes a discovery worthy of a Nobel Prize. The discovery might help to overcome a wide range of problems in the world. But what should we do if there is no explanation at all of what the discovery is or how it works?
AI is especially powerful in domains that deal with highly complex systems. Biological phenomena, cellular responses, the self-organization of materials, and nonlinear chemical-reaction networks are all extremely difficult to understand theoretically, and have therefore relied heavily on simulation. But we are now entering a stage in which, given sufficient data, AI can build predictive models directly. This unsettles the traditional process in science, in which one first understands a complex target and then models it.
AlphaFold is perhaps the clearest example. In protein structure prediction, if the mapping between input and output can be learned with sufficient accuracy, the model can still be practically useful even if its internal mechanisms are not intuitively understandable to humans. A functional model can exist even when its causal logic cannot be written down explicitly.
This forces human scientists to confront deeper questions: what counts as understanding, and what counts as explanation? How much explainability should be demanded? At what point should successful prediction be regarded as scientific understanding? The answers to these questions are not self-evident. They are what we must now begin to explore.
Explainability is also tied to the social legitimacy of AI. The principles that have long guided science, such as reproducibility, transparency, and accountability, do not necessarily align with black box hypothesis generation by AI. This is one of the social challenges that AI for Science is likely to face in the near future. That is why AGIS places emphasis not only on model development but also on experimental infrastructure, including robotics and automation. We aim to acquire data, record experimental conditions systematically, and improve reproducibility as far as possible without dependence on manual work. The opacity of AI should be counterbalanced with rigorous experimentation.
The goal of AGIS is not to force science into a closed loop of automation. It is to create a new and open ecosystem of knowledge in which AI, supercomputers such as Fugaku, robotics, and human scientists all interact and shape one another. In such an ecosystem, the value of science cannot be measured simply by efficiency or productivity. Alongside the pursuit of truth, one must also ask what kind of value science brings to society.

As scientists, we stand at a turning point in this larger transformation. We must ask again what science is, what understanding is, and how both should be remade. AGIS is a leading-edge organization that has chosen to take on that challenge and that responsibility. It is a place where change happens.

The problem that comes before AI

At AGIS, AI for Science is often discussed in terms such as accelerating scientific research or automating the research process. But before we even stand at that entrance, I believe we need to ask a different question first. How prepared is scientific research itself for AI intervention?
My own work has been in materials science, especially polymeric materials. These materials, including plastics, rubber, and fibers, are used throughout society and industry. Yet from the perspective of data-driven research with AI, this is a field that remains markedly underdeveloped. Put simply, it is a field in which data does not yet exist in sufficient quantity or in forms suitable for the assumptions on which AI depends.
Polymer materials offer enormous freedom in chemical structure, and the processes of synthesis and evaluation often involve many stages. As a result, each experiment is costly, and the data that is obtained tends to remain fragmented across laboratories. Under these conditions, a culture of sharing data and accumulating it openly across the field has scarcely developed.
AI does not function without data. This is not a matter of technological immaturity but a fundamental constraint. In fields where AI has produced striking results, such as shogi, Go, or protein structure prediction exemplified by AlphaFold, there were already vast data infrastructures built up and shared over many years. Research on polymer materials is a field that has lacked those conditions from the outset.
The framework I have proposed, polymeromics, is an attempt to change this situation. Rather than trying to understand polymer materials one by one as isolated cases, polymeromics aims to treat structure, synthesis conditions, physical properties, and function in an integrated way, so that these materials can be understood systematically and comprehensively.
In practice, this means carrying out large-scale simulations that reproduce the properties of polymer materials in silico and organizing the results into a systematic database. It also means using experimental robots to automate real-world synthesis and evaluation, thereby generating experimental data continuously. By linking that data with computational results, the goal is to accumulate knowledge in forms that AI can use.
In the field of life sciences, genomics and proteomics have generated new insights by enabling researchers to work across vast bodies of molecular information. The same kind of readiness, the ability to handle a field comprehensively, will be essential in the world of polymer materials. Polymeromics is an effort to build that foundation.

The problem of scale in AI-driven science

As I have worked on polymeromics, I have repeatedly encountered a series of challenges that lie between AI-driven science and more traditional scientific practice. I have tried not to treat them simply as defects, but rather as problems that will inevitably surface and therefore should be confronted early.
One of them is the problem of scale. By scale, I mean our sense of how much data, and how many trials, are required before scientifically meaningful discussion or prediction becomes possible.
In conventional experimental science on polymer materials, studies based on tens to hundreds of data points have often been sufficient for hypothesis testing and theory building. Researchers have read trends from limited data and used them to guide the next experiment or the next theoretical step. That sense of scale has shaped the development of the field.
But once the goal becomes to build general predictive models with AI, the premise changes substantially. Machine learning is not primarily a method for understanding individual cases. It is a method for learning the statistical structure embedded in data as a whole. For that reason, stable performance often requires at least hundreds of thousands of data points, and in some cases millions. In the longer term, one must even consider data sets on the scale of billions or tens of billions. That is not an exaggeration; it is a realistic requirement implied by the statistical character of current machine-learning methods.
Such a difference in scale, a difference of orders of magnitude, may produce a deep discontinuity between researchers’ intuitions and existing research systems. It is often assumed that once automated experimental platforms or robots are introduced, large quantities of data will follow naturally. In reality, things are not so simple. In many cases, even automated experiments produce only tens to hundreds of data points. Automation alone does not bring research to the scale that AI assumes.
So how can research move towards a scale that is meaningful for AI? What matters is not simply the apparatus itself, but a fundamental rethinking of experimental design. Which conditions should be explored, and over what range? Which properties should be prioritized for measurement? Decisions of that kind strongly shape both the quantity and the quality of the data that emerges.
This is not something that can be reduced to straightforward automation. It remains an area in which researchers’ experience and judgement are deeply involved. Indeed, it may be better understood as a domain in which the real question is how to combine human insight with AI’s computational power.
The work at AGIS is also a process of trial and error aimed at addressing these issues. The challenge is how to move from established research cultures and methodologies towards forms of research adapted to the age of AI, while still respecting what came before. Can a field remain sustainable while continuing to generate data that is scientifically meaningful over time? I believe that continuing to confront that question is one of AGIS’s central roles.

Doing science together in the places AI cannot reach

In research on polymer materials, the space of materials to be explored is, for all practical purposes, effectively infinite. No matter how advanced computational resources or experimental platforms become, there will be no day, at least not within the next several decades, when that space is filled in with data. In that sense, polymer materials research contains many regions that AI cannot readily reach.
The essence of science lies in exploring the unknown, in entering territory that has not yet been charted. And because that territory is unexplored, no prepared data exists there in advance. Without data, AI cannot intervene directly. This limitation is unlikely to disappear in any essential way, no matter how much the technology advances. AI is good at learning patterns from existing data. It cannot, by itself, step into places where no data yet exists.
As a result, the sciences in which AI is easiest to deploy may become skewed towards domains where data is relatively abundant. That risks introducing a bias into how science itself is pursued. For that reason too, AI for Science should not begin from the assumption that AI will solve everything.
What matters more is to start by recognizing that there will always be domains beyond AI’s reach, and then to design the division of labor between human researchers and AI accordingly. How far should tasks be entrusted to AI, and where must humans take over? Deliberately reconsidering that boundary is likely to become one of the defining demands of scientific research in the future.
In that sense, the idea of foundation models, as advanced at AGIS, is important because it asks how far human exploration can be supported while taking AI’s limits seriously. The idea is to prepare models that learn general knowledge from very large data sets, and then adapt them to specific domains in which only relatively small amounts of data can be obtained. This can be understood as an attempt to build a bridge into data-sparse areas that lie beyond AI’s easy reach.

I do not think AGIS’s role is to present a single correct answer to AI for Science. The issues involved, data, computation, experimentation, and even research institutions and research culture, are deeply entangled. What matters most is that AGIS serves as a place to test, one by one in real research settings, where to begin, and how to proceed.

Tackling science’s unresolved problems

Twentieth-century science produced many powerful theoretical tools for understanding nature. Relativity, quantum mechanics, and statistical mechanics made it possible to describe phenomena that appear intuitively complex in the language of mathematics, turning them into objects that could be predicted and controlled. These theories achieved many of their successes in what are known as equilibrium systems, systems in which interactions with the outside world are limited and the state does not change dramatically over time.
An equilibrium system is, for example, a system in a stable condition, such as a sealed container that has been left long enough for temperature and pressure to become uniform throughout. In such systems, the behavior of many components can be averaged, allowing the whole to be expressed in terms of a small number of variables. That, in turn, made it possible to draw relatively clear correspondences between theory and experiment.
But there were also domains that twentieth-century science could not fully address. These are the classes of problems broadly described as non-equilibrium systems and complex systems. Non-equilibrium systems are those in which energy or matter flows in and out, so that the state of the system continues to change. Living systems are a typical example. So too are climate change, brain activity, and social systems. These are also often described as complex systems.
In such systems, the difficulty is not only that they contain a vast number of components, in other words, a very large number of degrees of freedom. It is also that the components do not behave uniformly. Individual elements possess different states and histories, their own “individuality,” and the behavior of the whole emerges from the accumulation of their interactions. Simply averaging the behavior of cells, or of people, is not enough to explain the dynamics of life or society.
For that reason, simple averaging and approximation are much less effective in non-equilibrium systems, and the boundary conditions are not easily defined. The number of variables needed to represent the state of the system grows explosively, making it difficult to capture such systems in compact mathematical expressions of the kind that human cognition can readily handle. These high-dimensional, non-equilibrium systems are among the unresolved problems that twentieth-century science has handed on to the next generation.
This is precisely where AI for Science is likely to have its greatest impact. That does not mean AI is all-powerful or capable of solving every problem that has resisted previous approaches. What matters is that computational tools have finally become available at realistic scales that can handle many variables at once and model complex correlations while losing as little information as possible.
Nor is AI something that replaces existing mathematics or theory. AI does not take anything away from science. Instead, it is beginning to provide a new foothold in areas that previous theory and computation could not easily reach.

AI as the automation of intellectual labor

The phrase AI for Science often leads people to imagine a technology that will replace scientists themselves. That may be one aspect of it. But more importantly, it is also an attempt to define a new form of science, one that can enter domains that have resisted formalization and systematization.
Looking back at the twentieth century, the Industrial Revolution was above all a set of technologies for automating human physical labor. Steam engines and mechanization removed constraints on human strength and speed and dramatically expanded the scale and pace of production. In a similar way, AI in the twenty-first century may eventually come to be seen as a technology for automating intellectual labor. By intellectual labor, I mean not only tasks such as judgement, search and optimization, which contain some degree of regularity and repetition, but also thought processes in which creativity itself is at stake, such as the generation of new hypotheses.
What is the most concentrated site of creative trial and error in human activity? Alongside art, it is surely scientific research. For that reason, science is one of the fields most likely to adopt this new technology at the leading edge, to be transformed by it, and to benefit from it most deeply.

The essence of science lies where data is produced

Modern AI depends heavily on data and computational resources. That structure helps explain why large technology companies, which can collect vast amounts of information from the Internet and use it for training, have gained such a strong advantage.
But when thinking about AI for Science, one must remember that data is not only something that already exists. It must also be made. In experimental science, data typically comes into being only through carefully designed protocols of experiment and observation. By protocol here I mean the entire set of arrangements that define an experiment, including the conditions, measurement procedures, instruments, and treatment of error.
In that sense, the essence of science in AI for Science lies at the site where data is produced. What matters is not only the quantity or format of the data, but also the setting in which it was generated and the procedures through which it came into being.
At RIKEN, for example, advanced experimental techniques have been accumulated over many years, including organoid production, materials synthesis, and precise physical measurement, methods that can only be carried out in specific laboratory environments. These protocols are not merely difficult to reproduce. They are deeply embedded with researchers’ judgement and skilled practice, and can only be generated in advanced research settings.
If such experimental processes can be automated and linked with computation, then it may become possible to generate forms of scientific knowledge that AI companies alone could not easily reproduce.
The version of AI for Science envisioned at AGIS also differs from AI whose main purpose is to train on large volumes of pre-existing data. Its distinctive feature is that it does not stop at asking what can be learned from available data, but also asks what data should be generated next, and intervenes in that process. This approach is generally known as active learning. In active learning, AI proposes the next experimental conditions to test on the basis of hypotheses or uncertainty, automated systems such as robots carry them out, and the results are then fed back into the learning process.
What is being learned here is not only a static body of accumulated data. The very process through which data continues to be generated becomes part of learning itself. Put differently, the methodology of scientific research itself is being updated together with AI.
In this cycle, leadership does not necessarily belong to the model with the largest scale or the highest numerical accuracy. What matters is the ability to design and stably operate the setting in which particular kinds of data can be generated under particular conditions. That is why we are placing such emphasis on building autonomous robotic laboratories.

Humans still need human science

So what kind of science comes after AI for Science? Our world is an extraordinarily complex system made up of countless interacting elementary particles, unfolding across time and distance far outside ordinary human intuition. The human brain and senses have been shaped over hundreds of millions of years of biological evolution to help us survive in everyday life. Evolution did not require us to intuitively grasp timescales such as nanoseconds (one billionth of a second) or distances such as parsecs (about 3.26 light years). There is therefore no guarantee that our brains are naturally equipped to understand the objects studied by the natural sciences, whose history is only a few hundred years old.
And yet we try to simplify the world using expressions such as mathematics, language, and models, compressing it into forms of explanation that we can accept as intelligible. That activity is science. Science is what happens when human beings, with human bodies and brains, try to understand this complex world. If the human being is removed as the agent of science, there may still be mathematics, but there is no longer science. If that were not so, then the most complete understanding of the universe would simply be a full listing of the positions and states of all the elementary particles that compose it. But that is not how we ordinarily think. Science is the process of compressing information, through the human body and brain, into theories, equations and papers that human beings can understand.
At the same time, once one begins to think seriously about AI for Science, it becomes possible to imagine a future in which human science and AI science diverge. I have argued that understanding cannot exist without assuming a finite body and a finite cognitive structure. That is true for AI as well. Humans and AI process information in very different ways. If so, then some divergence between human scientific understanding and the kind of “scientific understanding” AI develops in order to improve its own performance may, in a sense, be inevitable.

Would such a divergence necessarily be a problem for human science? There will be many ways of thinking about that question, but I suspect it marks a major fork in the road for the future of science. At AGIS, we want to think about that divergence positively. In the long run, I believe the tension between human science and “AI science” could expand, rather than diminish, the possibilities of human science.

The divide in AI access is becoming a divide in scientific capability

AGIS’s mission under AI for Science is to change the situation in which only a limited number of researchers can make full use of AI, and instead to build it as an open research infrastructure. Put differently, the aim is to reinvent scientific research through AI.
We often speak of accelerating science. By that, we mean shortening the research cycle and increasing the number of innovations that can be reached. Attempts to accelerate science have of course been made before, through larger research budgets and greater investment in talent. But we have already entered an era in which those measures alone are no longer enough. Today, the keys to accelerating science are computational resources and AI.
Scientific progress can no longer be sustained solely by isolated achievements from a few exceptional researchers. Hypotheses are generated, tested, revised, and connected to the next question. When sufficient computational resources are allocated to that iterative process, and when AI, simulation, and automated labs are deeply integrated into it, science can begin to move beyond the limits imposed by the speed of human thought and the scale of human labor. We are approaching a point at which science can no longer be accelerated without such a foundation.
Already, disparities in access to AI are beginning to create an invisible divide in research practice. In the United States and Europe, the construction of research infrastructure built around the use of AI is advancing as a matter of national strategy, and research styles in which AI supports everything from hypothesis generation to experimental design and data analysis are spreading rapidly. Access to sufficient computational resources and AI changes the number of research cycles that can be run, and this produces decisive differences in the time required to reach results.
Some researchers around the world have already begun to work in environments in which explorations and validations that once took months can now be advanced in days or weeks. Others, working in the same fields, are at risk of being left behind by these changes. In time, this divide will become visible not only at the level of individual researchers but also in differences in scientific capability between institutions and even between nations. What we seek to achieve through AI for Science is to correct that divide, so that all researchers can benefit from AI and scientific research as a whole can be accelerated at the national level.

Towards a computing infrastructure that every researcher can fully use

At RIKEN, work is already underway within the TRIP Initiative to build the computational infrastructure needed for AI for Science. Positioned on an extension of that effort is the supercomputers Fugaku and its enhancements, as well as FugakuNEXT.
The goal of the effort is to provide an environment in which researchers can pursue science with AI without having to remain constantly aware of the computer itself. Experiment and computation, theory and data, human and AI should not remain separated, but should be connected as one continuous research process. That is the mission of enhancement of Fugaku, continuing on to FugakuNEXT as an infrastructure.
The original role played by the current supercomputer Fugaku was to demonstrate what advanced computing power itself could achieve. For scientific research that had long developed around experiment and observation, it showed through concrete results that computation could reproduce phenomena, deepen understanding, and even anticipate outcomes. Its achievement lay in demonstrating that computation is not merely a supporting tool, but can become a core driver of scientific progress.
The role of enhancing Fugaku with an AI for Science supercomputer, continuing on to FugakuNEXT is distinctly different. FugakuNEXT is conceived as something that goes beyond the traditional idea of a supercomputer as a device for executing large-scale calculations at high speed. It is designed on the assumption that AI and numerical simulation will be tightly integrated, and that the entire cycle of scientific research, including hypothesis generation and testing, code generation, and the automation of physical validation, will be made dramatically faster and more precise.
In that sense, FugakuNEXT is not merely a next-generation version of Fugaku. It is positioned as infrastructure intended to change the very methods of science. A researcher forms a hypothesis, tests it, interprets the results, and moves on to the next question. Within that cycle, AI is integrated naturally, and the speed and precision with which computation and experiment move back and forth are raised by orders of magnitude. FugakuNEXT is being designed as the foundation that supports that cycle.
And our mission is to make those computational resources available to all researchers. In other words, we want to accelerate science by building a computing infrastructure that every researcher can fully put to use.

Open science through open models

Through AI for Science, and through FugakuNEXT, we want to realize a form of open science based on open models. By open, we do not simply mean that results or code are made public. We mean a new form of science in which knowledge is shared in ways that allow anyone to participate, verify, and reuse it.
At present, many of the most advanced AI models are increasingly becoming black boxes. From the outside, it is often difficult to see what data they were trained on or through what procedures their results were produced. Such models are powerful in terms of convenience and performance, but they pose fundamental problems if they are to serve as the basis of science, which places high value on reproducibility and sharing. What matters in science is not only the result itself, but whether one can examine why that result was reached and whether that process is shared.
What AGIS is trying to build is an open computational infrastructure that can overcome this problem. The processes of designing, training, validating, and improving AI models should be operated openly, alongside the necessary computational resources. Data, algorithms, and computational processes should not be fragmented from one another, but maintained in a state in which they can be followed as scientific procedures and verified by third parties. That is what we mean by an open model.
If this kind of open science can be realized, the pace of research itself can be increased. If models and knowledge are shared on a common computational foundation, then work carried out through trial and error in one region does not need to be repeated from scratch elsewhere. Science can then begin to operate not as a collection of fragmented efforts, but as a continuous and collaborative process of updating shared knowledge.
Of course, openness also brings challenges, including those associated with dual use. But the era in which safety could be preserved only through closed and centralized systems has already come to an end. The history by which open-source software became established as social infrastructure makes that clear. It is increasingly understood that transparency, and scrutiny by many eyes, offers the best path for reconciling scientific reliability with social safety.
If Fugaku showed what computation can do, and FugakuNEXT asks how computation and AI can accelerate science, then open models represent the form of shared knowledge that lies beyond them. The aim is not to confine science to a small group of specialists, but to open it to society as verifiable knowledge. Building that new foundation is the task entrusted to AI for Science and FugakuNEXT.
Japanese supercomputers have long been highly regarded around the world. What Japan is now being asked to do is not to keep that reputation confined within its own borders. It is to contribute computational resources, data, and know-how within the global ecosystem, and to take an active role in building a shared scientific foundation. And within that effort, we must show clearly what contributions Japan has made. That, perhaps, is one vision of what it means to be a science-and-technology nation in the coming era.