Can MACE Potentials Accurately Describe Magnetism and Phase Stability in Fe-Ni Alloys? A Systematic Benchmark

This study demonstrates that a system-specific MACE potential trained on spin-polarized DFT data for disordered Fe-Ni structures significantly outperforms existing foundation models in predicting structural, elastic, and finite-temperature properties, though it still struggles to accurately capture magnetic collapse effects governing the bcc-to-hcp phase transition.

The Gist

Imagine you are trying to build a perfect digital twin of a complex machine made of iron and nickel. This machine is special because its behavior changes drastically depending on how much nickel you mix in, how hot it gets, and how much pressure you squeeze it with. Scientists call this an Fe–Ni alloy, and it's the kind of material found in everything from car parts to the very core of the Earth.

To simulate this machine on a computer, scientists need a "rulebook" called a potential. This rulebook tells the computer how every single atom should move and interact.

Here is what this paper did, explained simply:

1. The Problem: The "Generic" Rulebooks Didn't Work

Scientists already had some "foundation" rulebooks (called MACE foundation models) that were trained on huge, general datasets of many different materials. Think of these like a general encyclopedia: they know a little bit about everything.

However, the authors suspected these general rulebooks weren't detailed enough for the specific, tricky physics of iron-nickel alloys. Iron and nickel are "magnetic" and their atoms are messy and disordered. A general encyclopedia might miss the specific quirks of this particular alloy, especially when it comes to magnetism and how the material shrinks or expands under pressure.

2. The Solution: A Custom-Built "Specialized Manual"

Instead of using the general encyclopedia, the team built a custom rulebook (called MACE-sqs) specifically for iron-nickel.

• How they built it: They didn't just look at perfect, neat crystals. They used a technique called SQS (Special Quasirandom Structures). Imagine a bowl of M&Ms. A perfect crystal is like M&Ms arranged in a perfect grid. A real alloy is like a bowl where the colors are mixed randomly. The SQS method creates a digital bowl that perfectly mimics that random mix, capturing the "chaos" of real life.
• The Training: They fed this custom model data from high-precision quantum physics calculations (DFT) specifically for these random mixes. They taught it about energy, forces, magnetism, and how the atoms stretch and squeeze.

3. The Test: The "Exam"

The team put both the General Rulebooks and their Custom Manual through a series of rigorous tests to see which one could predict reality better.

• Test A: Squeezing the Material (Equation of State): They simulated squeezing the metal to see how much its volume shrinks.
  • Result: The Custom Manual was the most accurate. It matched real-world experiments almost perfectly. The General Rulebooks were often too "stiff" or too "squishy," getting the volume wrong.
• Test B: Stretching and Bending (Elasticity): They checked how the metal responds to stress.
  • Result: Again, the Custom Manual won. It correctly predicted how the metal gets harder or softer as you change the amount of nickel. The General Rulebooks missed some of the subtle, non-linear changes, especially in the "Invar" region (a specific mix of iron and nickel famous for not expanding when heated).
• Test C: The Phase Switch (BCC to HCP): Under extreme pressure (like deep inside the Earth), iron changes its internal structure from a cube shape (BCC) to a hexagon shape (HCP).
  • Result: This is where things got tricky. The Custom Manual predicted the pressure needed for pure iron to switch shapes reasonably well (closer to reality than the others). However, when they added nickel, all the models failed. They all predicted that adding nickel makes the switch happen at higher pressure, but experiments show it actually happens at lower pressure.
  • Why? The paper suggests the models are missing a specific "secret sauce": how the magnetism of the atoms collapses under high pressure. The models couldn't fully capture how nickel changes this magnetic collapse.

4. The Heat Test (Thermal Expansion)

They also tested how the metal expands when heated.

• Result: The Custom Manual did a great job predicting how the metal expands at normal temperatures. However, like all the models, it struggled a bit with the "Invar" effect (where the metal barely expands at all) and at very high temperatures where the magnetic order gets messy. This is because the model was trained on "frozen" magnetic states and didn't explicitly learn how to handle the chaotic "spin" of atoms at high heat.

The Bottom Line

Think of the General Rulebooks as a Swiss Army Knife: useful for many things, but not the best tool for any single specific job.

The Custom Manual (MACE-sqs) is like a specialized surgeon's scalpel. For the specific job of simulating iron-nickel alloys, it is far more accurate. It correctly predicts how the material behaves under pressure, how it stretches, and how it expands with heat.

The Catch: Even the best custom manual has a blind spot. It still doesn't fully understand what happens when you squeeze the material so hard that its magnetism collapses and it switches crystal structures. The authors conclude that to fix this, they need to teach the model even more about high-pressure magnetism and the hexagonal crystal structure, which they didn't include in the initial training.

In short: They built a better, more accurate digital twin for iron-nickel alloys by training it on messy, real-world-like data, but they still need to teach it a few more lessons about extreme pressure and magnetism.

Technical Summary

Technical Summary: Can MACE Potentials Accurately Describe Magnetism and Phase Stability in Fe–Ni Alloys?

Problem Statement
Iron–nickel (Fe–Ni) alloys serve as a canonical system for studying composition-dependent magnetism, magnetoelastic effects (such as Invar behavior), and phase stability under extreme conditions, including those relevant to planetary cores. Accurately modeling these systems is challenging because their properties are governed by a complex interplay between crystal structure, chemical disorder, and magnetic moments. While machine-learned interatomic potentials (MLIPs) offer a pathway to bridge the gap between the accuracy of density functional theory (DFT) and the computational efficiency required for large-scale simulations, the reliability of existing foundation models for chemically disordered magnetic transition-metal alloys remains uncertain. Specifically, it is unclear whether general-purpose models can capture the subtle magnetoelastic coupling and magnetic collapse phenomena that dictate phase transitions in Fe–Ni alloys.

Methodology
The authors present a systematic benchmark comparing several MACE foundation models against a system-specific model, MACE–sqs, trained on a targeted dataset.

• Reference Data: The reference dataset consists of spin-polarized PBE DFT calculations performed on chemically disordered Special Quasirandom Structures (SQS). These structures were generated to minimize short-range order (SRO) parameters, ensuring a representation of ideal random alloys across various compositions, crystal structures (bcc and fcc), and deformations (volumetric and shear).
• Model Training: The MACE–sqs model was trained exclusively on this SQS-based dataset (6,389 structures) to reproduce energies and forces. Crucially, the training utilized standard PBE without Hubbard $U$ corrections, motivated by the itinerant nature of Fe and Ni 3d states and the difficulty of defining a transferable $U$ across varying local environments.
• Comparative Models: The study evaluates MACE–sqs against foundation models trained on broader datasets, including those incorporating Hubbard $U$ corrections (e.g., MACE–mpa0, MACE–omat) and different exchange-correlation functionals (e.g., MACE–matpes with PBE and r2SCAN).
• Validation: The models were tested on equations of state (EOS), equilibrium volumes, elastic constants, magnetic moments, pressure-induced bcc-to-hcp phase transitions, and finite-temperature lattice expansion.

Key Contributions

1. System-Specific Training Strategy: The paper demonstrates that training MLIPs on chemically disordered SQS configurations, rather than relying solely on broad foundation datasets, significantly improves accuracy for Fe–Ni alloys.
2. Avoidance of DFT+U Artifacts: The study highlights that using PBE without Hubbard corrections as a reference level avoids the systematic distortions in equilibrium volumes and compressibility often introduced by $U$-corrected potentials in itinerant ferromagnets.
3. Comprehensive Benchmarking: The work provides a rigorous assessment of MACE potentials across structural, elastic, magnetic, and thermodynamic properties, identifying specific strengths and failure modes regarding phase stability under high pressure.

Results

• Accuracy: The MACE–sqs model achieves validation errors of 2.0 meV/atom for energies and 24.3 meV/Å for forces. It demonstrates the most consistent agreement with DFT and experimental data for equations of state, equilibrium volumes, and elastic constants across both bcc and fcc phases.
• Elastic Properties: MACE–sqs accurately reproduces the compositional trends of elastic constants ($C_{11}, C_{12}, C_{44}$) and derived moduli, including the non-monotonic behavior near the Invar composition, which foundation models often fail to capture or smear out.
• Finite-Temperature Expansion: The model successfully predicts thermal expansion coefficients at 300 K for various compositions. However, deviations at higher temperatures and near Invar compositions are attributed to the absence of explicit spin-disorder and spin-lattice coupling in the fixed-spin training framework.
• Phase Stability Limitations: A critical finding is the failure of all tested models (including MACE–sqs) to correctly predict the trend of the bcc-to-hcp transition pressure with increasing nickel content. While MACE–sqs predicts a pure-Fe transition pressure (15.7 GPa) closer to experiment than the foundation models, all models incorrectly predict an increase in transition pressure with Ni content, whereas experiments show a decrease.
• Root Cause of Failure: The authors attribute this discrepancy to the models' inability to fully capture high-pressure magnetic collapse and composition-dependent magnetoelastic effects, likely due to the lack of hcp training configurations under magnetic collapse conditions and the use of a collinear fixed-spin framework.

Significance and Claims
The paper concludes that targeted SQS-based training substantially improves the accuracy of MACE potentials for describing the structural, elastic, and finite-temperature properties of chemically disordered bcc and fcc Fe–Ni alloys within validated composition and pressure ranges. MACE–sqs is identified as a useful starting point for large-scale simulations of these systems.

However, the authors modestly claim that the current models are insufficient for predicting phase stability under conditions involving high-pressure magnetic collapse. The inability to reproduce the experimentally observed decrease in bcc-to-hcp transition pressure with nickel content indicates that "phase stability under magnetic collapse remains a key limitation for future model development." The work suggests that bridging this gap requires reference data that explicitly samples high-pressure magnetic collapse, hcp environments, and composition-dependent disorder effects, rather than relying solely on system-specific training of existing structural domains.