Machine Learning Helps Japan Create Stable Magnetic Quasicrystals

The story begins with a contradiction made solid. Inside a metallic alloy, atoms form an orderly pattern that never repeats. At low temperature, magnetic moments across that nonrepeating landscape align into a ferromagnet. And the route to the alloy did not begin with a metallurgist guessing one composition at a time. It began with a classifier reading the record of what generations of metallurgists had already made.

Researchers led by Professor Ryuji Tamura and Dr. Farid Labib at Tokyo University of Science report three bulk, annealable ferromagnetic icosahedral quasicrystals made without the rapid-quenching process on which earlier ferromagnetic quasicrystals depended. A machine-learning phase classifier screened 675 five-element alloy systems and helped direct the team toward gold–copper–aluminum–indium alloys containing one of three rare-earth elements: gadolinium, terbium or dysprosium. The candidates still had to survive the furnace, diffraction measurements, calorimetry and magnetometry. Three did.

The result, published online in the Journal of the American Chemical Society on July 7, 2026, matters less as an instant commercial magnet than as a clean new laboratory. Earlier ferromagnetic quasicrystals were rapidly frozen, metastable and difficult to improve by heating: annealing could turn them into periodic “approximant” crystals. The new materials can be arc-melted in bulk and annealed for long periods at 723 kelvin—about 450 degrees Celsius—while retaining and improving their quasiperiodic order. That makes the strange magnetism measurable rather than merely observable.

What the Japanese Team Actually Made

The word “stable” requires care. It does not mean that these alloys remain quasicrystalline at every temperature or that they are ready for a motor. It means the quasicrystalline phase can be produced without an ultrafast freeze and can withstand prolonged annealing in the tested regime. Tokyo University of Science reports that the materials remain stable at and below 723 K; above that point they transform toward a related periodic approximant. At 723 K, extended heat treatment sharpened the X-ray diffraction peaks, evidence that atomic disorder was being reduced and quasiperiodic coherence improved rather than destroyed.

That reverses the central weakness of the 2021 and 2023 ferromagnetic quasicrystals. Rapid quenching cools a molten alloy so quickly that atoms can be trapped in a phase they would not choose if given time to rearrange. It is a powerful discovery technique, but it tends to produce thin ribbons, small samples, strain, compositional inhomogeneity and competing phases. Annealing ordinarily heals a metal. In the earlier ferromagnetic quasicrystals it could erase the very phase under investigation.

All three showed clear ferromagnetic transitions. Their heat-capacity curves displayed pronounced lambda-shaped anomalies at the transition temperatures, while magnetization measurements separated the relatively free-to-rotate Gd moments from the directionally constrained Tb and Dy moments. The compounds share the same broad quasiperiodic architecture, yet changing a single rare-earth species changes the fluctuations that dominate the transition. That is the scientific prize: a controlled comparison on a lattice that does not repeat.

A Crystal That Refuses to Repeat

A conventional crystal is ordered because a unit cell repeats by translation: move by the right distance and the atomic environment returns. An amorphous solid such as ordinary glass lacks that long-range positional order. A quasicrystal occupies a third category. Its atoms are ordered over long distances, producing sharp diffraction spots, but there is no single repeating unit cell that tiles the material by translation.

The familiar two-dimensional analogy is a Penrose tiling. Two kinds of rhombus can cover a plane with local rules and recognizable motifs, yet the full pattern never repeats periodically. An icosahedral quasicrystal is a three-dimensional metallic counterpart with rotational symmetries—fivefold axes among them—that conventional periodic crystallography once excluded. A regular pentagon cannot tile a flat floor by itself; similarly, exact icosahedral symmetry cannot coexist with an ordinary three-dimensional translational lattice.

That did not prevent nature from arranging atoms that way. On April 8, 1982, Dan Shechtman, then examining rapidly solidified aluminum–manganese alloy by electron microscopy, recorded a diffraction pattern with tenfold symmetry. The observation contradicted the accepted definition of a crystal and met fierce resistance. Shechtman and collaborators published the result in 1984. The International Union of Crystallography later broadened the definition of a crystal around essentially discrete diffraction, and the Nobel Committee awarded Shechtman the 2011 chemistry prize “for the discovery of quasicrystals.”

Japan entered the history almost immediately and decisively. In 1987, An-Pang Tsai, Akihisa Inoue and Tsuyoshi Masumoto at Tohoku University reported a well-ordered, thermodynamically stable Al–Cu–Fe icosahedral quasicrystal. Stable high-quality phases made it far harder to dismiss quasicrystals as artifacts of a violent cooling process. Tsai and collaborators went on to discover families of stable quasicrystals and, in 2000, the first stable binary quasicrystal in cadmium and ytterbium. The local multi-shell atomic clusters associated with this lineage are now known as Tsai-type clusters.

That Japanese history is not a decorative preface to the 2026 result. The new magnetic compounds are Tsai-type hypermaterials, and the team’s predictive variables, candidate database and magnetic design ideas are products of the experimental family tree that began with those stable alloys.

Why Magnetism Was the Hard Part

Ferromagnetism is collective. Below a Curie temperature, many atomic magnetic moments choose a common orientation, producing spontaneous magnetization. In iron the phenomenon lives on a periodic lattice whose translational symmetry gives theory a powerful simplifying tool. In a quasicrystal every site can be locally familiar while the environment never repeats in the ordinary way. The question is not just whether individual rare-earth atoms carry moments; it is whether interactions can coordinate them across an aperiodic structure.

For decades, magnetic quasicrystals usually froze into spin-glass-like states. Moments became stuck in disordered directions rather than developing a single coherent long-range order. Chemical disorder, competing interactions and geometric frustration all offered explanations. The experimental record left open a deeper suspicion: perhaps quasiperiodicity itself frustrated conventional magnetic order.

Approximant crystals became the bridge. They contain atomic clusters closely related to those in a quasicrystal but arrange the clusters periodically, allowing cleaner structural analysis and conventional theoretical tools. In 2010, researchers established long-range antiferromagnetic order in a Cd₆Tb approximant. Over the following decade, composition changes in gold-based approximants showed that the balance between ferromagnetism, antiferromagnetism and glassy freezing could be tuned.

A major control knob was the valence-electron concentration, written e/a: the average number of relevant valence electrons per atom. In these alloys, localized 4f moments on rare-earth atoms communicate through conduction electrons via the oscillatory Ruderman–Kittel–Kasuya–Yosida, or RKKY, interaction. Changing e/a shifts the electronic environment and can change whether the effective coupling at a given separation favors parallel or antiparallel alignment. The new alloys cluster at e/a values of 1.74 to 1.77, close to the ferromagnetic region established in earlier Tsai-type work.

In 2021 Tamura’s team and collaborators reported the first long-range ferromagnetic order in real icosahedral quasicrystals, Au–Ga–Gd at 23 K and Au–Ga–Tb at 16 K. Neutron diffraction provided magnetic Bragg evidence. In 2023, the group added a high-phase-purity, composition-tunable Au–Ga–Dy quasicrystal and showed a relationship between magnetism and e/a. But those samples still relied on rapid quenching. The 2026 work closes that process gap.

The Machine Learned from a Small, Human-Built World

Materials discovery is an awkward machine-learning problem. The possible space of compositions and processing histories is immense, while reliable positive examples are scarce and expensive. A database records what scientists chose to attempt, not all that is physically possible. Failed syntheses are less likely to be published. A model can therefore learn the habits and blind spots of a field along with its physical regularities.

The Japanese group’s approach developed in stages. In 2021, researchers from the Institute of Statistical Mathematics, the University of Tokyo and Tokyo University of Science trained a random-forest classifier on the compositions of 80 known quasicrystals, 78 approximants and 10,090 ordinary crystals. It classified quasicrystal, approximant and “other” phases with an overall reported accuracy near 0.714, while identifying ordinary crystals almost perfectly. Interpreting the model revealed rules involving atomic size, electronegativity and electron concentration; it even recovered the familiar Hume–Rothery tendency near e/a = 1.8 for many aluminum-based quasicrystals.

In 2023, a refined binary classifier predicted whether a composition would form a thermally stable quasicrystal with reported accuracy above 95 percent. Guided by its ranking, researchers synthesized Al₆₅Ni₂₀Os₁₅, Al₇₈Ir₁₇Mn₅ and Al₇₈Ir₁₇Fe₅—described as the first quasicrystals discovered by a machine-learning algorithm. The claim did not mean the machine performed the experiment. It meant a learned pattern chose candidates that human synthesis then confirmed.

The data foundation widened in 2024 with HYPOD-X, an open database assembled by the Institute of Statistical Mathematics, Tokyo University of Science and the National Institute for Materials Science. Experts manually or semi-automatically extracted and checked compositions, structure types, heat-treatment conditions, digitized phase diagrams, and temperature-dependent thermal, electrical and magnetic properties from papers and books. Its three datasets—composition, phase diagrams and properties—turned a fragmented literature into a form that algorithms could use.

For the 2026 search, a phase classifier using HYPOD-X and other databases generated 675 quinary systems. The Au–Cu–Al–In–R family rose among the leading candidates. But the model was one instrument in a chain of reasoning. Researchers still selected rare-earth members with known magnetic moments, prepared alloys by arc melting, controlled annealing, checked phase identity by diffraction, and measured heat capacity and magnetization down to cryogenic temperatures.

This is a strong model for responsible “AI for science.” The algorithm proposes where scarce experimental attention may be most valuable. It does not get the last word.

Three Rare Earths, Two Kinds of Critical Behavior

Near a continuous magnetic phase transition, quantities such as magnetization, susceptibility and heat capacity change according to power laws. Their critical exponents tell physicists how fluctuations propagate and whether apparently different systems belong to the same universality class. Periodic crystals have supplied the textbook cases. A high-quality quasicrystal asks what remains universal when translational symmetry is removed.

The Gd alloy behaved as a relatively isotropic Heisenberg-like system. Gd³⁺ has essentially no orbital angular momentum, so its moment is less strongly pinned to a preferred local direction. It saturated at a comparatively low field of about 0.6 tesla, and its critical behavior departed markedly from the mean-field values.

Tb³⁺ and Dy³⁺, by contrast, experience strong crystal-electric-field anisotropy. Their moments are constrained by local atomic environments and did not saturate even at the study’s maximum 7-tesla field. Both displayed critical parameters close to mean-field predictions. The team interprets the contrast as the result of stronger spin fluctuations in Gd and fluctuation suppression by anisotropy in Tb and Dy, interacting with the shared quasiperiodic structure.

“Mean-field-like” should not be read as a photograph of literally infinite-range forces. Mean-field theory replaces the detailed influence of neighboring spins with an average effective field. Agreement with its exponents is a clue about the effective range and suppression of fluctuations, not a license to ignore microscopic structure. Indeed, the authors’ own 2026 review of Tsai-type magnetism identifies universality classes and microscopic ordering mechanisms as open questions.

The achievement is that the comparison can now be made on annealed material with sharper quasiperiodic coherence. Structural imperfection is no longer as easy an explanation for every anomaly. The quasicrystal itself becomes an experimental variable.

Why “Stable” Is More Important Than “AI”

The machine-learning label attracts attention, but annealability changes the science. A bulk sample can support repeated measurements, better statistics and a wider range of probes. Heat treatment can improve structural order. Researchers can compare samples processed for different durations, examine defects and phason strain, and eventually attempt larger grains or single-quasicrystal growth. Neutron scattering, muon spin rotation, resonant X-ray methods and local probes all benefit from more material and cleaner phases.

Stability also permits the difficult separation of structure from disorder. If a rapidly quenched ribbon contains approximant inclusions, compositional gradients and frozen strain, a magnetic anomaly may come from several causes. An annealed high-coherence phase lets theory confront quasiperiodicity more directly.

Yet the stability window must not be inflated into a device claim. The reported Curie temperatures are between 9.7 and 28.3 K—far below room temperature. The alloys contain substantial gold and scarce rare-earth elements. The study did not report coercivity, electrical losses, mechanical durability, manufacturability, cycling life or device performance. A material can be a superb platform for fundamental physics and a poor commercial permanent magnet.

Tokyo University of Science points to possible long-term relevance for sensing, energy conversion and information processing. Those are research directions, not product forecasts. The near-term output is knowledge: a new way to test magnetic theory where periodicity is absent.

Japan’s Quiet Infrastructure for Materials Intelligence

The discovery joins two Japanese strengths that are often narrated separately. One is patient alloy science: phase diagrams, arc melting, diffraction, low-temperature measurement and an institutional lineage extending from Tohoku’s stable quasicrystals to Tokyo University of Science’s magnetic hypermaterials. The other is materials informatics: converting experimental literature into validated datasets and interpretable predictions.

HYPOD-X is particularly important because fashionable models cannot learn from data that do not exist in structured form. Its records were not simply scraped and trusted. Experts reviewed compositions and fabrication conditions; phase boundaries were digitized; property curves were extracted from figures. That labor is easy to overlook because it does not look like artificial intelligence. It is the reason the prediction has scientific memory.

The project also crossed organizational borders. The 2026 author list spans Tokyo University of Science, the University of Tokyo, Aoyama Gakuin University, the Institute of Statistical Mathematics, RIKEN and the National Institute for Fusion Science. Funding came from Japan Society for the Promotion of Science grants and the Japan Science and Technology Agency’s CREST program. The collaboration brought together metallurgy, low-temperature physics, crystallography and data science.

There is a broader policy lesson. Materials AI is most credible where public infrastructure preserves negative and positive experiments, processing histories, measurement uncertainty and machine-readable phase data. A dramatic discovery can begin with a classifier, but the durable national asset is the database and the community willing to maintain it.

The Questions the New Quasicrystals Make Possible

The next experiments are more interesting than an immediate hunt for a product. Can the new phases be grown as large single grains? What magnetic structure will neutron diffraction resolve? How do RKKY interactions propagate across a quasiperiodic network of rare-earth icosahedra? Do phason degrees of freedom—rearrangements peculiar to quasicrystals—couple to spins? Does the Gd compound define a new critical universality class, or will better measurements place it within an existing one?

The whirling, noncoplanar spin arrangements found in related approximant crystals suggest another frontier. Quasiperiodic magnets may host unusual textures and transport responses, potentially including anomalous or topological Hall effects. But the route from a suggestive cluster model to a measured bulk effect is long. Stable samples make that route traversable.

Materials design poses equally practical questions. Can electron concentration and local anisotropy be tuned without expensive gold? Can transition temperatures rise while the quasicrystalline phase remains stable? Can antiferromagnetic, ferromagnetic and glassy states be moved predictably through composition? Can active learning choose the next experiment using both success and failure rather than only the published winners?

These questions reveal why the 675-system screen is not the end of a search. It is the beginning of a feedback loop. Every synthesis, failed or successful, can improve the map—if the result is recorded with enough detail to learn from.

From the Impossible Pattern to a Testable Material

Quasicrystals have repeatedly moved from impossibility to instrument. Shechtman’s diffraction pattern forced crystallography to accept ordered matter without periodic repetition. Tsai’s stable alloys showed the phase could be more than a rapidly frozen accident. Magnetic approximants revealed design rules. The 2021 quasicrystals proved long-range ferromagnetism could coexist with quasiperiodicity. The 2026 alloys make that coexistence stable enough to interrogate.

Machine learning deserves credit, but not mythology. It worked because decades of difficult experiments had been organized into data; because physicists knew which variables mattered; because metallurgists could make the candidates; and because measurements were allowed to reject the model’s suggestions. The scientific intelligence belonged to the whole chain.

For Japan, that chain may be the larger achievement. It links a 1980s materials tradition to open data and contemporary computation without discarding the furnace, the diffraction pattern or the skeptical measurement. The algorithm did not abolish trial and error. It made the next error more informative and the next trial less blind.

A quasicrystal is ordered without repeating. The discovery process now has a similar character: history supplies motifs, machine learning recombines them, and experiment decides whether a genuinely new pattern has appeared. In three small pieces of gold-rich alloy, it has.