Surface segregation in high-entropy alloys from alchemical machine learning

Arslan Mazitov, Maximilian A. Springer, Nataliya Lopanitsyna, Guillaume Fraux, Sandip De, Michele Ceriotti

TL;DR

Researchers have developed a machine learning potential capable of simultaneously modeling 25 transition metals to predict surface segregation in high-entropy alloys (HEAs). By using "alchemical" compression to map 25 elements onto a smaller feature space, the model efficiently predicts which elements migrate to an alloy's surface—a critical factor for designing next-generation catalysts.

What Problem Does This Paper Address?

High-entropy alloys (HEAs) are "metallic soups" containing five or more elements in near-equal proportions. They are exciting for catalysis because their complex surfaces offer a vast "design space" for chemical reactions. However, modeling them is a nightmare:

  1. Scale: To represent a disordered alloy, you need large simulation cells that make traditional ab initio (DFT) calculations prohibitively slow.
  2. Complexity: Most machine learning (ML) models scale exponentially with the number of elements. A model for 25 elements usually requires a massive amount of data and memory.
  3. Surface Segregation: The elements at the surface of a catalyst are rarely the same as the bulk concentration. Predicting this "segregation" is vital because the surface composition dictates catalytic performance.

The authors sought to create a general-purpose ML model that could handle 25 elements and accurately describe surface phenomena without needing an astronomical amount of new training data.

Methods

The authors extended their previous HEA25-4-NN model. The core innovation is alchemical contraction:

  • Feature Compression: Instead of treating all 25 elements as distinct entities (which would lead to a combinatorial explosion of parameters), the model uses a coupling matrix to project the 25 elements into a 4-dimensional "pseudo-element" space. This captures the "chemical personality" of elements based on their position in the periodic table.
  • Incremental Learning: They started with a model trained exclusively on bulk crystalline data. They then "fine-tuned" it by adding a relatively small number of "defective" configurations (surfaces and melts)—less than 20% of the original dataset size.
  • Sampling: To find the equilibrium surface state, they used Replica Exchange Molecular Dynamics (REMD) combined with Monte Carlo (MC) atom swaps. This allows the simulation to overcome high energy barriers and find the most stable arrangement of atoms.

Datasets

The study utilizes the HEA25S dataset, which includes:

  • Subset O: ~25,000 bulk structures (fcc/bcc) from previous work.
  • Subset A: ~2,600 on-lattice surface slabs with various Miller indices.
  • Subsets B & C: ~2,000 snapshots of bulk liquids and thermally equilibrated surfaces generated via MD.
  • Subset D: 500 structures specifically mimicking a 25-element Cantor-style alloy surface.

Key Insights and Ideas

  • Transferability of "Alchemy": The alchemical weights learned from bulk crystals remained almost unchanged when optimized for surfaces. This suggests that the underlying chemical relationships between transition metals are robust across different structural environments.
  • Size Drives Segregation: The model "rediscovered" a classic physical trend: elements with larger atomic radii (like Ag or Au) tend to segregate to the surface to relieve the tensile stress inherent in the bulk lattice.
  • The Danger of Static Models: The authors show that "static" segregation energy calculations (common in the literature) can be misleading. For example, Chromium appears to deplete from the surface in static tests but actually accumulates in the sub-surface layer when full thermodynamic sampling is used.
  • Metastability in Quinary Alloys: In the study of IrFeCoNiCu, the model revealed the alloy is actually thermodynamically unstable and tends to phase-separate into Ni-Cu and Ir-Fe-Co regions, providing a new perspective on experimental observations of these materials.

Key Results

  • High Accuracy: The final model (HEA25S-4-NN) achieved energy errors between 10–16 meV/atom across all subsets, including complex surface and molten states.
  • Validation: The model's predictions for the CoCrFeMnNi (Cantor alloy) matched explicit DFT calculations with less than 10% error in segregation enthalpy.
  • General Propensity Map: By simulating a 25-element alloy, the authors produced a "surface excess" map, identifying elements at the edges of the d-block (like Ag, Pd, Ti) as having the strongest surface affinity, while central elements (like Fe, Mn) tend to stay in the bulk.

Limitations and Future Work

  • Magnetism: The current model is trained on non-spin-polarized DFT data. While this is a consistent choice for a broad 25-element model, it neglects magnetic effects that are important for Fe, Co, and Ni at low temperatures.
  • Environmental Effects: These simulations were performed in a vacuum. Real-world catalysts operate in the presence of oxygen, water, or reactants, which can drastically alter segregation patterns (e.g., "pulling" certain metals to the surface via oxidation).
  • Kinetics: The study focuses on thermodynamic equilibrium. In many HEAs, "sluggish diffusion" means the surface state seen in experiments might be a trapped metastable state rather than the global minimum predicted here.

The authors suggest that future iterations could incorporate magnetic descriptors and explicit gas-phase molecules to model "in operando" catalytic conditions.

Summary updated

Abstract

Abstract High-entropy alloys (HEAs), containing several metallic elements in near-equimolar proportions, have long been of interest for their unique mechanical properties. More recently, they have emerged as a promising platform for the development of novel heterogeneous catalysts, because of the large design space, and the synergistic effects between their components. In this work we use a machine-learning potential that can model simultaneously up to 25 transition metals to study the tendency of different elements to segregate at the surface of a HEA. We use as a starting point a potential that was previously developed using exclusively crystalline bulk phases, and show that, thanks to the physically-inspired functional form of the model, adding a much smaller number of defective configurations makes it capable of describing surface phenomena. We then present several computational studies of surface segregation, including both a simulation of a 25-element alloy, that provides a rough estimate of the relative surface propensity of the various elements, and targeted studies of CoCrFeMnNi and IrFeCoNiCu, which provide further validation of the model, and insights to guide the modeling and design of alloys for heterogeneous catalysis.
BibTeX
@article{mazitov2024surface,
  author = {Arslan Mazitov and Maximilian A. Springer and Nataliya Lopanitsyna and Guillaume Fraux and Sandip De and Michele Ceriotti},
  title = {{Surface segregation in high-entropy alloys from alchemical machine learning}},
  journal = {Journal of Physics Materials},
  year = {2024},
  doi = {10.1088/2515-7639/ad2983},
}

Loading related content...

Figure