Micro-Physical AI: The Atomic-Scale Frontier of Physical AI and the Future of Materials Discovery

What is Physical AI?

Current LLMs (Large Language Models) learn about the world from text. Through this, they incorporate the knowledge accumulated by humans so far and are already having a major impact on the world.

However, to borrow Plato's words, text is merely a "shadow" of reality. Text only records a portion of what humans have been able to verbalize, and the overwhelming majority of phenomena occurring in the physical world have not yet been turned into words. Furthermore, there are many phenomena that cannot be described in words in the first place. Much of the physical world lies behind the conclusions summarized in papers, books, and reports as a vast "undocumented reality."

Physical AI is an AI that structures and makes manageable this undocumented physical world itself as knowledge. Long before the emergence of Physical AI, we have been working toward "making the real world itself computable." The phrase "making computable" includes not only making things manageable on computers but also taking action in the real world. In this context, the highlighted Physical AI largely aligns with our vision. This is the major challenge we will tackle over the next 10 years.

The Other Invisible Physical AI

When people hear the term "Physical AI," many probably associate it with humanoid robots or autonomous vehicles. Here, I would like to introduce another important field of Physical AI: the micro-scale physical world of atoms and molecules. Let us call this "Micro-Physical AI." What I contend here is that within Physical AI, Micro-Physical AI also holds the key to significantly shaping the future of society.
​

Our world is built upon a stack of various scales. The human scale, from centimeters to meters, is a visible and familiar realm, but inside it lies the structure of materials, the materials themselves, and even deeper within, the behavior of atoms and electrons.

The size of an atom is approximately 0.1 nanometers. If we were to expand 1 meter to the size of the Earth, an atom would finally become about a few centimeters in size. Thinking about atoms at our scale is like looking at a pebble on the scale of the Earth. Despite this, the extremely micro-level behavior at the atomic level determines the ultimate properties of substances and materials.

For example, the performance of semiconductors that support AI is determined by how efficiently the internal materials can conduct electricity. The performance of batteries essential for electric vehicles and robots also varies greatly depending on the properties of their constituent materials. Carbon dioxide absorbents for achieving a sustainable society are likewise determined by the properties of their molecular structures.₂​

The "Materials Barrier" Binding Society

Furthermore, to solve critical modern challenges, we inevitably hit the problem of materials.

All-solid-state batteries are highly anticipated as next-generation batteries. If realized, they will enable safe, high-performance energy systems, but the key lies in an extremely micro-level phenomenon: how smoothly ions can move through a solid. Energy density, degradation, short-circuiting—all remaining challenges are atomic-scale problems.

The same applies to the semiconductors supporting AI. Current forecasts suggest that an additional power capacity on the scale of tens of gigawatts will be required to run the AI supporting global society. In this regard, the realization of AI semiconductors, data centers, and optoelectronic integration can significantly reduce this power consumption, making the evolution of AI practical. Beyond this, the key to realizing nuclear fusion power generation and quantum computers also lies in new materials.

From this, it can be said that future technology depends heavily on materials. It is no exaggeration to say that society today is bound by the constraints of the materials barrier.​​

The Scaling Law of Intelligence Confronts the Scaling Law of Reality

Materials science fundamentally grapples with the sheer complexity of the physical world. To create new materials, it is necessary to face and understand new phenomena that did not exist in the past. To achieve this, there is no choice but to engage in massive trial and error—namely, repeating experiments while facing resource constraints in the physical world, or conducting simulations based on physical laws.​​

On the other hand, with conventional simulations, increasing precision caused an explosion in calculation time, making it impossible to handle realistic scales.

Today, AI is changing this situation. For example, systems like Matlantis have enabled calculations that are orders of magnitude faster than conventional methods while maintaining high accuracy by learning the interactions between atoms. Similarly, in various fields, simulations are being dramatically accelerated by AI, and the prediction of reactions and material properties is becoming possible at a practical level.

On the other hand, strongly correlated electron systems, where electrons strongly influence each other—which is an important phenomenon determining the individuality of substances such as superconductivity and magnetism—behave in an extremely complex manner that even current AI cannot easily handle. Between atomic-level behavior and the emergence of macro-level material properties, there is a hierarchy of complexity beyond imagination. It is like looking at a few-centimeter sample and talking about the weather of the entire Earth.

While the realization of intelligence exceeding humans is often cited in the world to debate whether certain jobs will disappear, in the materials field, even if intelligence exceeding humans is realized, the challenges of materials development will not all be solved. No matter how much AI evolves, it will not become capable of predicting all physical phenomena in advance.​​

The complexity of the real world will remain indefinitely, and exploration will continue.

A Future Where Researchers and AI Agents Develop Materials Together

In this context, let us imagine a future realized by Micro-Physical AI.

In the field of software development, development utilizing AI agents is achieving unprecedented acceleration. However, it is difficult to bring AI agents directly into the field of materials development. This is because parallelizing and accelerating the experimental part is difficult. Unlike software development, which is self-contained within a virtual world, in Physical AI, experiments in the real world become the bottleneck.

Here, simulation plays a crucial role. By using dramatically accelerated simulations, AI agents can break free from the constraints of the physical world and advance unprecedented research within a parallelized and accelerated world. Of course, experiments are also indispensable. This part will be handled by automated experimental systems realized by macro-scale Physical AI.

For example, suppose a researcher wants to create higher-performance and safer battery materials. The vast majority of the process—where an AI agent investigates past papers and patents, compares existing materials using the company's historical data, forms hypotheses, runs simulations, and repeats experiments—will be automated. Consequently, parallelization and acceleration will be achieved on an unprecedented scale. Researchers will review summaries of the process or parts deemed important, and occasionally intervene directly.

Furthermore, humans and AI will engage in an iterative dialogue to consider future research directions. Based on past results, they will determine which goals to prioritize and what constraints to establish—human working hours will not be a bottleneck. Humans will spend most of their time critiquing the continuously emerging results and designing research themes and environments. Such activities are called "harness engineering," which involves organizing existing and private data into a format easily accessible by AI, as well as setting up simulations and experimental environments. Notably, the ongoing activity of maintaining these setups will likely be supported by AI as well.

Countless groups of agents will continuously run a vast number of simulations and automated experiments 24 hours a day, 365 days a year. If unexpected results occur, they will be shared with researchers for additional validation. Furthermore, automated experiments utilizing robotics will be constantly planned, ensuring that the verification of hypotheses continues ceaselessly.

What We Aim For

Materials science currently defines the upper limits of technological progress. Batteries, semiconductors, nuclear fusion, quantum computers—nearly all technologies that influence the future of humanity are blocked by the materials barrier. Pushing back this barrier through the fusion of AI, computation, and experimentation is the domain we are challenging.

Micro-Physical AI is a collaborative effort between humans and AI to understand reality itself in an invisible, micro world, and it is a challenge that will shape the future.