Nov. 17Research published in 2026
3 frequencies100 kHz · 1 MHz · 10 MHz
1988Early FINEMET alloy research
2021Fugaku entered shared operation
Scientific qualification: this study maps an exchange-derived energy barrier during simulated reversal. It does not directly measure heat dissipation or demonstrate lower electricity use in a working device.

A hidden bottleneck in electrification

Every time a power converter changes voltage or switches current, its magnetic components must respond. Transformers and inductors do much of that work out of sight. Some energy is lost in the process, and those losses matter when electric vehicles, data centers and industrial equipment depend on ever more demanding power electronics.

On November 17, researchers from Tokyo University of Science and collaborating universities announced an unusual way to investigate where part of that difficulty originates. They combined large-scale simulations on Japan’s Fugaku supercomputer with topology-based, explainable machine learning to identify selected magnetic-domain walls associated with a rising energy barrier during high-frequency magnetization reversal.

The open-access study appeared the same day in NPG Asia Materials. Its achievement is not a measured reduction in electricity consumption, nor a new commercial magnetic alloy. It is a method for relating intricate microscopic patterns to a physically interpretable measure of the resistance encountered when magnetization changes direction.

Why soft magnets matter to hard-working electronics

The term soft magnetic does not describe something pliable to the touch. It refers to materials whose magnetization can change relatively easily under an applied magnetic field. That makes them useful in components that must repeatedly store, release and redirect magnetic energy. Permanent magnets play a different role.

Inside a magnetic solid, neighboring regions called domains have magnetization pointing in broadly similar directions. Domain walls separate those regions. When a changing field drives the material, the walls move and deform, some domains advance, and others retreat. The resulting choreography is microscopic and extraordinarily complex.

Engineers can measure core loss in a component and estimate contributions from hysteresis and eddy currents. But a bulk number does not necessarily tell them which local wall patterns impede reversal most strongly. Connecting local form with measurable performance is a longstanding materials-science problem.

Three frequencies, one simulated material

The team modeled a FINEMET-type iron-based nanocrystalline soft magnet using micromagnetic simulations performed with Fugaku. They examined responses at 100 kilohertz, 1 megahertz and 10 megahertz—conditions that span a demanding high-frequency regime for magnetic components.

The simulations tracked how domain patterns evolved as magnetization reversed. At higher excitation frequencies, domains tended to divide into finer structures, and an energy barrier associated with reversal became more pronounced near the coercive region.

It matters that these were simulations of an idealized material. The team did not film an actual operating core at all three frequencies or report a converter-level energy saving. Fugaku, made available for shared scientific use in March 2021, supplied the computational scale needed to generate the evolving magnetic configurations for analysis.

The mathematics of shape becomes a scientific instrument

A domain image looks like an irregular network of regions and boundaries. Ordinary pixel statistics can miss how its parts connect. The researchers instead used persistent homology, a method from topology that can encode features such as connectivity and holes across changing scales.

They then used principal component analysis to compress the structural descriptors and connect them to an image-derived measure of magnetic wall energy. The result was an interpretable energy landscape: a way of placing changes in the microstructure beside changes in a physically meaningful quantity.

This is what distinguishes the work from simply training a black-box algorithm to predict a number. Its design allows researchers to trace important model features back toward the spatial structures that produced them.

Not every domain wall does the same job

When the researchers mapped important features back onto the simulated domain images, the contributions were not spread evenly across every wall. They were concentrated around the advancing tips of reversed domains and selected small wall segments.

That observation suggests an internal division of labor. Some structures appear to hinder reversal by contributing to an energy barrier; other wall configurations can enable the process to move forward. Broadly similar-looking boundaries may therefore play markedly different physical roles.

The analysis points toward a more targeted question for materials designers: could composition, processing or microstructure be adjusted so that magnetization reverses with less resistance? It does not yet establish that such adjustments will succeed in a manufactured component.

The crucial difference between an energy barrier and lost heat

One caveat is essential to interpreting the headline. The study linked magnetic patterns to an exchange-derived structural energy barrier. That is not the same thing as directly measuring dissipated heat or total core loss in a working device.

Actual power loss depends on several mechanisms, among them hysteresis, eddy-current effects, geometry and operating conditions. The authors acknowledge that a fuller treatment could incorporate eddy currents, magnetostriction and magnetostatic energy. A feature important to an exchange-energy barrier cannot automatically be assigned a measured fraction of real-world heating.

The researchers have not supplied a verified percentage reduction in EV energy use, data-center electricity demand or transformer loss. That limitation does not diminish the computational finding; it defines the scientific work still required to connect the finding to engineering gains.

A Japanese materials legacy dating to 1988

The material family in the simulation has a history closely tied to Japanese industrial research. In 1988, Yoshizawa, Oguma and Yamauchi at Hitachi Metals reported a new class of iron-based nanocrystalline soft magnetic alloys. A 1989 Japanese technical paper further described the development of what became known as FINEMET.

The key idea was to create exceedingly fine crystals within a remaining amorphous matrix through controlled heat treatment. This combination helped give the material useful soft magnetic characteristics and comparatively low losses, contributing to decades of development in magnetic cores and other electrical components.

The latest study is therefore not the invention of FINEMET. It brings newer tools—high-performance computing, geometric data analysis and explainable machine learning—to a material tradition in which Japan has long played an important role.

Why the stakes extend from EVs to data centers

Electric vehicles need power conversion for motor drives and auxiliary systems. Data centers rely on multiple stages of conversion to deliver tightly controlled power to electronics. Magnetic cores appear within many of those systems, and their behavior can constrain efficiency, thermal management and component size.

Switching at higher frequencies can allow smaller passive components, but can also make magnetic losses harder to manage. Better material selection is therefore not merely a matter of choosing a higher operating frequency. The magnetics, semiconductor switches, thermal design and control strategy must be considered together.

The team’s approach may eventually help researchers look beyond bulk loss curves and diagnose which structural features change under demanding waveforms. Whether that translates into lower manufacturing cost, higher efficiency or improved reliability remains to be tested.

An example of AI designed to explain, not just predict

AI for science is sometimes presented as a contest to automate discovery. This work illustrates a more restrained and arguably more useful possibility: computational tools can organize a massive body of physical data and make a subtle pattern intelligible to scientists.

The collaboration spans Tokyo University of Science, the University of Tsukuba, Okayama University, Kyoto University, Tohoku University and Johannes Gutenberg University Mainz. Its expertise combines magnetic physics, topology, computation and materials informatics. Masato Kotsugi of Tokyo University of Science is the study’s senior researcher; Kosuke Nishioka is the first-listed author.

The team identifies experimental techniques such as synchrotron-based measurements and time-resolved Kerr microscopy as promising next sources of data. A model that can make sense of laboratory images, rather than simulations alone, would be more directly useful to materials developers.

What would make this discovery industrially consequential?

The next test is reproducibility across real materials. Grain sizes, mechanical stress, alloy composition, manufacturing defects and temperature can all alter magnetic behavior. A correlation found in a controlled simulation might weaken or change under the messy conditions of a production line.

Then comes the engineering experiment: adjust a material or process in response to the model, measure the resulting core loss and demonstrate that any improvement holds under representative operating conditions. Only after that can manufacturers make credible claims about watts saved in a power converter or battery miles gained in a vehicle.

For now, the discovery is a map rather than a finished machine. It marks parts of a microscopic magnetic landscape that deserve closer scrutiny—and gives researchers a new way to ask why energy barriers emerge in the first place.

Primary sources and verification limits

This independent report relies on the November 17 announcements from the participating universities, the article’s accessible bibliographic record, and historical primary sources. We did not interview the researchers, repeat simulations or measure hardware. The journal full text was access-restricted during review; detailed methodological claims were cross-checked against the university explanations. No commercial device-level saving is confirmed.

  1. Tokyo University of Science: Japanese primary announcement, Nov. 17, 2026
  2. NPG Asia Materials: peer-reviewed research article, DOI 10.1038/s41427-026-00683-2
  3. Tokyo University of Science: English research account and limitations
  4. Okayama University: joint research announcement
  5. Kyoto University: collaboration and research team
  6. Tohoku University: institutional research statement
  7. RIKEN: Fugaku supercomputer entered shared operation in 2021
  8. Yoshizawa and Yamauchi: 1989 historical paper on FINEMET
  9. Yoshizawa, Oguma and Yamauchi: landmark 1988 nanocrystalline alloy paper