🔬 materials science

Coercivity-Aware Machine Learning Discovery of Rare-Earth-Free Soft Magnetic Alloys with First-Principles Magnetic Benchmarks

This paper presents a coercivity-aware machine learning framework that integrates experimental data screening with first-principles magnetic benchmarks to identify and prioritize rare-earth-free Fe-Co-rich soft magnetic alloys, while highlighting the critical importance of composition-grouped validation and the distinct influence of microstructure on coercivity prediction.

Original authors: Avik Mahata, David Priefer

Published 2026-10-05

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Avik Mahata, David Priefer

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

Modern life runs on electricity, but moving that power efficiently requires more than just wires; it demands materials that can guide magnetic fields with ease. These are called soft magnetic materials, and they are the silent workhorses inside electric motors, transformers, and the charging systems for our vehicles. For these devices to work well, the material must be able to switch its magnetic direction quickly without losing energy, a property known as low resistance to change, or coercivity. It also needs to stay magnetic even when it gets hot, which is measured by a specific temperature threshold called the Curie temperature. For decades, engineers have relied on rare-earth elements to make these materials perform, but those elements are expensive and their supply is unreliable. This has created a urgent need to find new alloys made from common metals like iron, cobalt, and nickel that can do the same job without the rare-earth cost.

The challenge in finding these new materials is that their performance depends on a complex mix of their chemical recipe and how they are manufactured. Two pieces of metal with the exact same chemical makeup can behave very differently if one is cooled quickly and the other slowly. This makes it difficult to predict which new combination of elements will work best just by looking at a list of ingredients. A team of researchers has tackled this problem by building a new kind of computer model that learns from past experiments while being careful not to overfit. Instead of simply testing the model on data it has already seen, they forced it to predict the properties of entirely new chemical recipes it had never encountered before. This approach revealed that while it is possible to predict how hot a material can get before losing its magnetism, predicting how easily it can switch directions is much harder because that property is so sensitive to the manufacturing process.

To solve this, the researchers gathered a massive collection of experimental data from scientific literature, compiling over 1,700 records of magnetic materials. They used this data to train a machine learning system, but they set a strict rule: the system had to learn from groups of known chemical recipes and then be tested on completely different groups it had never seen. This prevented the computer from simply memorizing answers. They tested two main goals: finding alloys that stay magnetic at high temperatures and those that switch magnetic directions with very little resistance. The computer learned that for the high-temperature goal, knowing the exact mix of elements was enough to make a good guess. However, for the switching resistance, the computer struggled because the data showed that the same chemical mix could produce very different results depending on how the metal was processed. This confirmed that a computer cannot predict the switching resistance of a new alloy just by looking at its ingredients; it needs to know how the metal was made.

Despite this difficulty, the team used their model to screen 20,000 potential new recipes made from iron, cobalt, nickel, manganese, aluminum, and silicon. They applied a safety net to their predictions, calculating a range of possible outcomes to ensure they didn't pick a candidate that might fail in reality. This careful filtering narrowed the list down to a small group of promising candidates that are rich in iron and cobalt. To be sure these computer predictions were grounded in physical reality, the researchers then ran independent simulations based on the fundamental laws of physics. These simulations confirmed that the iron-cobalt structures they identified do indeed possess strong magnetic properties and can withstand high temperatures, providing independent physical benchmarks for the computer's choice of candidates rather than a direct validation of the specific recipes.

The result is a shortlist of eight specific chemical recipes that serve as hypotheses for further investigation. The top candidate, for instance, is a mix containing roughly 47 percent iron, 24 percent cobalt, and 18 percent nickel, with small amounts of aluminum and silicon. The computer predicts this alloy could stay magnetic up to 940 Kelvin, with a very low resistance to switching, though the uncertainty in the prediction means the actual resistance could be higher. The researchers emphasize that these are not finished products but rather the most promising starting points for further investigation. By combining a smart computer search with independent physics checks, they have provided a clear path forward for developing affordable, rare-earth-free magnets that could power the next generation of electric vehicles and energy systems.

Technical Summary: Coercivity-Aware Machine Learning Discovery of Rare-Earth-Free Soft Magnetic Alloys

Problem Statement
The increasing electrification of transportation and industrial systems demands soft magnetic materials that combine efficient flux transport with thermal stability, while avoiding reliance on critical rare-earth elements. Developing these materials is complicated by the coupled dependence of magnetic properties on composition, processing, and microstructure. Specifically, while intrinsic properties like saturation magnetization and Curie temperature ( $T_C$ ) are composition-dependent, coercivity ( $H_c$ ) is heavily governed by microstructural features (grain size, defects) and processing history. Existing predictive models often struggle to distinguish between intrinsic compositional effects and extrinsic processing effects, and standard machine learning (ML) evaluation protocols (random partitioning) often fail to assess a model's ability to generalize to entirely new, unseen compositions due to data leakage from identical compositions appearing in both training and testing sets.

Methodology
The authors developed a "coercivity-aware" machine learning workflow designed to screen rare-earth-free soft magnetic alloys (specifically the Fe–Co–Ni–Mn–Al–Si system) using coupled design targets: Curie temperature and coercivity.

Dataset and Descriptors: The study utilized a subset of the Northeast Materials Database (NEMAD), containing 1,736 records with both $T_C$ and $H_c$ data. Three descriptor sets were evaluated:
- Elemental fractions (77 features) plus valence electron concentration and arithmetic density.
- Composition-weighted elemental statistics (146 features) based on atomic properties (e.g., radius, electronegativity, magnetic moment).
- A combined set of 225 features.
- Coercivity was modeled as the dimensionless logarithm ( $\log_{10}(H_c/1 \text{ Oe})$ ).
Evaluation Protocol: To rigorously test transferability to unseen chemistries, the authors employed composition-grouped partitioning. Unlike random partitioning, this method ensures that no encoded composition appears in both the training and testing sets. The evaluation involved 20 repeated grouped partitions with nested model selection (choosing the best descriptor set and learner from Random Forest, Extra-Trees, and Gradient Boosting).
Uncertainty Quantification: Residual-based prediction intervals were constructed using validation set residuals to provide uncertainty bounds for predictions. Candidates were screened using these intervals alongside a distance-to-training filter (L1 distance in elemental fraction space) to ensure candidates were not excessively far from the training distribution.
Physical Benchmarks: To independently validate the ML screening results, the authors performed:
- First-Principles Calculations: Density Functional Theory (DFT) calculations on representative Fe–Co structures (bcc Fe, B2 FeCo, L1 $_2$ Fe $_3$ Co, L1 $_2$ FeCo $_3$ ) to assess magnetic polarization and magnetovolume effects.
- Finite-Temperature Simulations: Atomistic spin simulations (Monte Carlo) to track the loss of magnetic order and estimate transition scales for these benchmark systems.

Key Results

Model Performance: Composition-grouped evaluation yielded lower but more realistic predictive accuracy compared to random partitioning. The mean test $R^2$ values were 0.601 ± 0.080 for Curie temperature and 0.589 ± 0.069 for logarithmic coercivity. Random partitioning produced higher scores ( $R^2 \approx 0.74$ ) but included significant composition overlap, failing to measure true transferability.
Descriptor Impact: Combining elemental fractions with weighted elemental statistics improved $T_C$ prediction ( $R^2$ increased from 0.600 to 0.630). However, for coercivity, the combined set did not improve performance over elemental fractions alone ( $R^2 \approx 0.565$ vs. $0.587$), consistent with the strong dependence of $H_c$ on processing variables not captured by composition descriptors.
Screening Outcomes: From 20,000 sampled Fe–Co–Ni–Mn–Al–Si compositions:
- Only 11 candidates met the strict point-prediction criteria ( $T_C > 1000$ K, $H_c < 100$ Oe).
- None of these 11 satisfied the stricter interval criteria at these thresholds.
- At relaxed thresholds ( $T_C > 600$ K, $H_c < 100$ Oe), 127 candidates satisfied both the interval and distance filters. These candidates are predominantly Fe–Co-rich.
Physical Validation: First-principles calculations confirmed strong magnetic polarization and magnetovolume effects in Fe–Co structures. Finite-temperature spin simulations indicated that these benchmark Fe–Co systems retain magnetic order up to high temperatures (transition scales $T_{0.1}$ ranging from ~1110 K to 1320 K), providing independent physical support for the ML-identified Fe–Co-rich region.

Significance and Claims
The paper claims to establish a computational framework for prioritizing rare-earth-free soft-magnetic chemistries that explicitly distinguishes between composition-level screening and candidate-specific physical validation.

Methodological Contribution: The study demonstrates that composition-grouped evaluation is essential for assessing models intended to screen new alloy chemistries, as random partitioning artificially inflates performance metrics by leaking identical compositions.
Uncertainty-Aware Discovery: By integrating residual-based prediction intervals and distance filters, the framework narrows the candidate space while explicitly acknowledging predictive uncertainty, preventing the over-interpretation of point predictions.
Physical Consistency: The work does not claim to have discovered a specific, validated alloy. Instead, it provides a "coercivity-aware" workflow that identifies promising compositional regions (Fe–Co-rich) and validates them against independent physics-based benchmarks (DFT and spin simulations), distinguishing the statistical screening process from the physical verification of specific candidates.

The authors emphasize that the identified candidates are hypotheses for further structural, processing, and magnetic evaluation, rather than application-qualified soft magnets, as critical properties like hardness, ductility, and core loss were not included in the screening objectives.

Technical Summary: Coercivity-Aware Machine Learning Discovery of Rare-Earth-Free Soft Magnetic Alloys

More like this