How Explainable AI Revealed the Magnetic “Maze” Behind Energy Loss in Electric Motors

Electric motors rely on soft magnetic materials that constantly flip their magnetization as current changes direction. Each flip costs energy. A major part of that loss—known as iron loss or magnetic hysteresis loss—comes from microscopic magnetic domains inside the motor core rearranging themselves during magnetization reversal.

Researchers at Tokyo University of Science (TUS) developed a physics‑guided, explainable AI framework to analyze one of the most complicated forms of these structures: maze‑like magnetic domains. By combining topological data analysis with an entropy‑extended Ginzburg–Landau free‑energy model, the team mapped how these domains evolve, uncovered the energy barriers governing magnetization reversal, and clarified why certain materials waste energy during operation.

The Hidden Magnetic “Maze” in Motor Materials

Soft magnetic materials used in motor cores—such as nonoriented electrical steel—contain tiny regions called magnetic domains. Each domain has a uniform magnetization direction, but neighboring domains often point in different directions.

In some materials, these domains form complex zig‑zag networks known as maze domains. These structures can change dramatically with temperature and external magnetic fields, making them difficult to model with traditional physics tools.

Because magnetization reversal requires domains to shift, merge, or split, the detailed geometry of these domain patterns strongly influences how much energy is lost during each cycle of operation.

Turning Domain Images Into Quantitative Physics

To study these complex patterns, the researchers built a data pipeline that starts with microscope images of magnetic domains.

The key analytical step uses persistent homology, a technique from topological data analysis. Persistent homology extracts measurable features from complex spatial structures—such as loops, branching patterns, and connectivity—allowing maze‑like domain patterns to be represented as quantitative descriptors rather than just images.

This approach transforms chaotic domain maps into structured data that machine‑learning systems and physics models can analyze.

Extending the Ginzburg–Landau Energy Model With Entropy

Traditional Landau and Ginzburg–Landau models describe phase transitions and magnetization using energy terms tied to physical interactions. However, these models struggle to represent the enormous number of possible configurations in maze‑like domains.

The TUS team addressed this limitation by adding an entropy feature to the free‑energy framework, creating an entropy‑feature‑extended Ginzburg–Landau (eX‑GL) model.

This modification allows the model to represent not only energetic contributions but also the statistical complexity of domain configurations. In practice, the entropy term captures how many microscopic arrangements are possible within a given domain pattern.

By fitting topological features from persistent homology into this extended free‑energy landscape, the system can identify:

• energy barriers between domain configurations
• pathways of magnetization reversal
• temperature‑dependent transitions in domain morphology

The result is an explainable AI model that connects observable domain patterns to the underlying physical energy landscape governing them.

Revealing the Energy Barriers Behind Magnetization Reversal

Using the eX‑GL framework, the researchers analyzed how maze‑domain structures evolve as temperature changes and external magnetic fields drive reversal.

The model showed that complex domain patterns correspond to specific energy barriers in the magnetic free‑energy landscape. These barriers determine how easily domains can reorganize during magnetization reversal.

Instead of treating magnetization reversal as a simple energetic transition, the entropy‑extended model captures the competition between:

• magnetic interaction energies
• structural complexity of domain networks
• thermal effects that change domain stability

This combination explains why reversal in maze‑domain materials can be abrupt, temperature‑dependent, and highly dissipative.

Connecting Microscopic Domains to Macroscopic Energy Loss

One of the most important outcomes of the work is a direct link between microscopic domain structure and macroscopic magnetic hysteresis behavior.

The framework allows automated identification of the mechanisms responsible for energy loss in materials such as nonoriented electrical steel by analyzing how domain features map onto the free‑energy landscape during reversal.

Because hysteresis loss originates from irreversible changes in domain configuration, identifying the relevant energy barriers reveals where energy is dissipated in the cycle.

Why This Matters for Electric Vehicle Motors

Electric vehicles rely on high‑efficiency motors built from soft magnetic materials. During operation, the magnetic field in the motor core repeatedly flips direction, forcing domains to reorganize and causing iron loss.

By identifying the domain structures and temperature conditions that create large energy barriers, the new framework could help researchers and engineers:

• design steel microstructures that minimize dissipative domain configurations
• optimize processing conditions that control domain morphology
• predict how materials behave under real motor operating conditions

Improving the magnetic efficiency of these materials could reduce energy loss inside motors and contribute to higher overall efficiency in electric vehicles, although the current research does not quantify a specific efficiency gain.

A Physics‑Guided Path for Explainable AI in Materials Science

Beyond motor materials, the study demonstrates how explainable AI combined with physics models can analyze complex microstructures that were previously too chaotic to quantify.

Instead of treating machine learning as a black box, the TUS framework integrates interpretable features—derived from topology and thermodynamics—into a physically meaningful model. This approach provides both predictive capability and scientific insight into the mechanisms governing complex materials.

As materials science increasingly relies on data‑driven methods, frameworks like the entropy‑extended Ginzburg–Landau model offer a way to merge AI with established physical theory to uncover hidden mechanisms in microscopic systems.