🔬 materials science

Orbital fingerprinting of magnetism across the Mn-Ni-Ga Heusler ternary

This paper introduces orbital fingerprints derived from projected density of states as effective descriptors for machine-learning models to accurately predict magnetic ordering, moments, and phase diagrams across the Mn-Ni-Ga Heusler ternary system, overcoming the limitations of traditional composition-based rules.

Original authors: Felipe Hawthorne, Daniela A. Damasceno, Raphael M. Tromer, Ronaldo R. Pelá, Cristiano F. Woellner

Published 2026-09-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Felipe Hawthorne, Daniela A. Damasceno, Raphael M. Tromer, Ronaldo R. Pelá, Cristiano F. Woellner

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the world of materials science, some alloys are like chameleons, changing their behavior based on how they are made or what environment they find themselves in. Among these, a family of metals known as Heusler alloys stands out for their ability to switch between different magnetic states. These are not just simple magnets that stick to a refrigerator; they are complex mixtures of elements that can stretch, shrink, or change their internal order when exposed to magnetic fields or temperature shifts. This makes them valuable for creating advanced devices, from sensors that detect tiny movements to cooling systems that use magnetic fields instead of harmful gases. However, predicting exactly how a specific mixture will behave is notoriously difficult. While scientists have long relied on simple rules based on the ratio of ingredients to guess a material's properties, these rules often fail when the mixture is slightly off-balance or when the atoms are arranged in a disordered way. The magnetic state of these materials depends on a delicate dance of electrons, and even a tiny shift in the recipe can cause the entire system to flip from one magnetic personality to another.

A team of researchers set out to solve this puzzle by looking at a specific trio of elements: manganese, nickel, and gallium. They wanted to understand how the magnetic personality of this mixture changes across the entire range of possible recipes, not just the perfect, textbook versions. To do this, they turned to a powerful combination of computer simulations and artificial intelligence. Instead of trying to guess the outcome based on the list of ingredients alone, they taught a computer to read the "electronic fingerprint" of the material. This fingerprint is a detailed map of how electrons are distributed around the atoms, revealing subtle clues about how the atoms interact with one another. By training their models on hundreds of simulated structures, the researchers discovered that this electronic map contains the true key to predicting magnetic behavior, far surpassing the accuracy of traditional methods.

The researchers began by generating a vast library of 370 different atomic arrangements for the manganese-nickel-gallium system. These were not perfect, orderly crystals but rather realistic, slightly disordered structures that mimic the messy reality of actual materials. For each structure, they ran complex computer simulations to determine the lowest-energy state, effectively asking nature which magnetic configuration the atoms preferred to settle into. The results revealed a rich and competitive landscape. In some regions of the mixture, the atoms aligned perfectly in the same direction, creating a strong ferromagnet. In others, they arranged themselves in opposing patterns, canceling each other out to become non-magnetic, or forming complex, partially opposing structures known as ferrimagnets. The boundaries between these states were incredibly thin, often separated by energy differences so small they were barely measurable, making it nearly impossible to predict the outcome just by looking at the chemical formula.

To make sense of this complexity, the team introduced a new way of describing the materials using what they called orbital fingerprints. Imagine taking a snapshot of the energy levels of the electrons orbiting specific atoms, such as the manganese or nickel, and breaking that snapshot down into its most important features. These features included how far apart the energy levels were for electrons spinning in different directions and how many electrons were occupying specific energy bands. The researchers fed these fingerprints into machine-learning models, training them to recognize the patterns that led to each magnetic state. The results were striking. When the models relied only on the basic recipe—the percentage of manganese, nickel, and gallium—they could correctly guess the magnetic state about two-thirds of the time. However, once the models were given the electronic fingerprints, their accuracy jumped to nearly 87 percent.

The most significant finding was identifying exactly which part of the electronic fingerprint mattered the most. The models consistently pointed to a single feature: the energy gap between the electron states of manganese atoms. This gap, which represents the difference in energy between electrons spinning one way versus the other, acted as the primary switch determining whether the material would be magnetic and what kind of magnetism it would exhibit. This discovery confirmed that the underlying physics driving the material's behavior was indeed captured by these fingerprints. Furthermore, the researchers found that these fingerprints could also predict the strength of the magnetic moment and the degree of spin polarization, a property crucial for next-generation electronics, with a high degree of precision. In contrast, traditional methods based on counting valence electrons failed completely in this disordered environment, producing results that were worse than random guessing.

The study also highlighted the potential for discovering new materials with unique properties. By analyzing the energy differences between competing magnetic states, the team identified several specific compositions where the energy gap was vanishingly small. In these rare cases, a tiny external push, such as a magnetic field, could tip the balance and switch the material from one state to another. This sensitivity is the hallmark of the magnetocaloric effect, a phenomenon that could be harnessed to create highly efficient, eco-friendly cooling technologies. The researchers noted that while their simulations were powerful, they were limited to cubic structures and did not account for more complex, distorted shapes that might occur in real-world manufacturing. Nevertheless, the approach they developed offers a robust and generalizable tool. It bridges the gap between the slow, expensive process of calculating every detail from scratch and the fast but often inaccurate method of simple recipe-based guessing. By learning to read the electronic fingerprints, scientists now have a clearer path to designing the magnetic materials of the future, turning the chaotic competition of atoms into a predictable and tunable science.

Technical Summary: Orbital Fingerprinting of Magnetism across the Mn–Ni–Ga Heusler Ternary

Problem Statement
In Mn–Ni–Ga Heusler alloys, the competition between magnetic ground states (ferromagnetic, antiferromagnetic, ferrimagnetic, and non-magnetic) is governed by energy differences of only a few meV per atom. Consequently, the magnetic ground state cannot be reliably predicted from chemical composition alone, particularly in off-stoichiometric, binary, or elemental regimes where established rules like the Slater–Pauling rigid-band picture break down. While first-principles calculations (DFT) can resolve these states, they are computationally expensive, making large-scale screening of the ternary composition space impractical. Conversely, existing machine learning (ML) approaches either rely solely on composition (failing to distinguish magnetic orderings with identical stoichiometry) or require extensive training data specific to certain systems, limiting their applicability to broad ternary screening. There is a critical need for descriptors that explicitly encode electronic-structure details governing magnetism while remaining efficient enough for ML-based classification and regression.

Methodology
The authors constructed a dataset of 370 spin-polarized DFT calculations covering the full Mn–Ni–Ga ternary space. The structures were modeled as cubic special quasirandom structures (SQS) with supercell sizes ranging from 2 to 10 atoms to simulate compositional disorder, rather than ordered intermetallic phases. Calculations were performed using collinear spin-polarized DFT (SIESTA, PBEsol functional) without Hubbard U corrections, justified by the itinerant nature of the Mn 3d states.

To bridge the gap between electronic structure and ML, the authors introduced orbital-resolved projected density-of-states (PDOS) fingerprints as descriptors. These fingerprints were derived from the orbital- and spin-resolved DOS of Mn 3d, Ni 3d, Ga 4p, and Mn 4s states. The descriptor set included:

Composition-only features: Mole fractions, electronegativity spreads, and valence electron counts.
PDOS fingerprints (64 features): Spectral moments (band centers, bandwidths, third moments), Fermi-level PDOS, occupied-state counts, spin polarization, and exchange splitting ( $\Delta_{ex}$ ).
Structural descriptors: Lattice parameters, cell angles, volume, and stress.
Moment-derived descriptors: Used primarily to diagnose target leakage (as they encode the ground truth labels).

Four ML models (Gradient-Boosted Trees, Random Forests, Support Vector Machines, and k-Nearest Neighbors) were benchmarked using stratified 5-fold cross-validation and Leave-One-Composition-Out (LOCO) validation to test generalization to unseen stoichiometries.

Key Results

Magnetic Ordering Classification:
- Models trained solely on composition achieved an accuracy of 68.1%.
- Incorporating PDOS fingerprints increased accuracy to 86.8%.
- The top-ranked descriptor was the Mn 3d exchange splitting ( $\Delta_{ex}$ ), which aligns with first-principles findings that this quantity governs the separation of magnetic phases.
- LOCO validation yielded a mean accuracy of 76.0%, indicating that the PDOS–ordering relationship captures transferable physics rather than memorizing specific compositions.
- The inclusion of moment-derived descriptors (target leakage) pushed accuracy to 93.0%, quantifying the information gap between PDOS fingerprints and direct moment data.
Magnetic Moment Amplitude Regression:
- Composition-only models yielded $R^2 = 0.728$ .
- PDOS-based models achieved $R^2 = 0.898$ (MAE = 0.10 $\mu_B$ ).
- Contrary to the Slater–Pauling rule (which failed with $R^2 = -4.19$ for this dataset), the dominant feature for moment prediction was the electronegativity spread ( $\delta_{EN}$ ), suggesting that interspecies charge transfer plays a more significant role in modulating moment magnitude than valence electron count in off-stoichiometric systems.
Fermi-Level Spin Polarization:
- Composition-only models failed to predict spin polarization ( $R^2 \approx 0$ ).
- PDOS fingerprints enabled quantitative prediction with $R^2 = 0.937$ .
- Crucially, even when Fermi-level specific descriptors were withheld, the model retained $R^2 = 0.521$ , demonstrating that the band shape away from the Fermi level contains predictive information about the polarization at $E_F$ .
Magnetocaloric Candidates:
- The DFT dataset identified four compositions with energy differences ( $\Delta E$ ) between competing magnetic states below 5 meV/atom (e.g., MnGa $_3$ , Mn $_9$ Ga), marking them as potential candidates for magnetocaloric refrigeration.

Significance and Claims
The paper claims that orbital-resolved PDOS fingerprints serve as an effective "systematic bridge" between first-principles electronic structure and machine learning predictions. The approach successfully captures the underlying physics of magnetic ordering—specifically exchange splitting and orbital hybridization—without requiring the atomic spin moments as input.

The authors position this method as a middle ground between fast but inaccurate composition-only screening and accurate but expensive full electronic-structure calculations. By requiring only a single self-consistent DFT calculation per structure to extract descriptors, the method enables efficient exploration of composition and ordering across a magnetic ternary system. The study highlights that while composition alone suffices for estimating magnetic moments in some contexts, it is insufficient for classifying magnetic ordering or predicting Fermi-level spin polarization, tasks where electronic-structure fingerprints are essential.

Limitations
The authors note two primary limitations:

The use of cubic SQS supercells excludes ordered site arrangements found in tetragonal phases (e.g., $L1_0$ MnGa, $D0_{22}$ Mn $_3$ Ga) and the associated antisite ferrimagnetism in Mn-rich compositions.
The collinear spin approximation and small supercell sizes (2–10 atoms) preclude the study of non-collinear spin textures and long-period magnetic orderings.

More like this