🔬 materials science

Data-Driven Prediction of NaCl-Type Entropy-Stabilized Oxide Compositions from First-Principles and Supervised Learning

This study establishes an efficient data-driven framework that combines density functional theory, special quasirandom structures, and supervised machine learning to accurately predict the stability and stabilization temperatures of equimolar quinary NaCl-type entropy-stabilized oxides, thereby enabling the rapid screening of thousands of potential compositions with minimal computational cost.

Original authors: Sebastien Junier, Celine Barreteau, David Berardan, Yann-Andrev Kerneur, Jean-Claude Crivello

Published 2026-07-07

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Sebastien Junier, Celine Barreteau, David Berardan, Yann-Andrev Kerneur, Jean-Claude Crivello

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

Imagine you are a chef trying to create the perfect "super-salad." You have a pantry with 16 different types of vegetables (the chemical elements). Your goal is to mix exactly five of them together in equal amounts to create a single, uniform dish that stays stable and doesn't separate into layers. In the world of materials science, this is called an Entropy-Stabilized Oxide (ESO).

However, with 16 vegetables, there are 4,368 possible five-vegetable combinations. Trying to cook every single one in a real kitchen (a laboratory) to see which ones work would take forever and cost a fortune.

This paper describes how the researchers built a super-smart digital kitchen to predict which combinations will work, saving them from having to cook them all physically.

Here is how they did it, broken down into simple steps:

1. The Recipe Book (The Database)

Before they could predict new salads, they needed to know how the individual ingredients behaved. They used powerful computer simulations (called DFT) to create a massive "recipe book" of known, stable oxide structures.

The Analogy: Think of this as testing every single vegetable on its own and in small pairs or triples to see how they react to heat. They used two different "tasting methods" (mathematical functions): a standard one (GGA) for speed and a more precise one (meta-GGA) for accuracy.

2. Simulating the Chaos (SQS)

Real entropy-stabilized oxides are messy. The atoms are jumbled randomly, like a salad where the lettuce, tomatoes, and cucumbers are thoroughly mixed. Standard computer models struggle with this randomness.

The Analogy: To handle the mess, the researchers used a trick called Special Quasirandom Structures (SQS). Imagine taking a small, manageable bowl of salad that looks perfectly random from every angle, even though it's just a small sample. They used these "mini-bowls" to simulate what happens when you mix all five ingredients together without actually needing a billion atoms in the computer.

3. The Crystal Ball (Machine Learning)

Even with the "mini-bowls," calculating all 4,368 combinations is too slow. So, they trained a computer brain (Machine Learning) to be a crystal ball.

The Analogy: They cooked about 420 of the 4,368 combinations in the digital kitchen and fed the results to the computer brain. They taught it to look at the "ingredients" (properties like atomic size and electrical charge) and guess the stability of the other 3,900+ combinations without cooking them.
The Result: A specific type of computer brain (a Multi-Layer Perceptron) became very good at guessing. It could predict the stability of a new mix with an error margin of about 4 kJ/mol. This is accurate enough to tell you if a salad will hold together or fall apart, even if it can't tell you the exact temperature down to the degree.

4. The Taste Test (Validation)

The researchers didn't just trust the computer; they went back to the real kitchen to test their predictions.

The Good News: The computer correctly identified known "winning" recipes that scientists had already discovered.
The Reality Check: When they tried to make some of the new predicted recipes, they found that some of them didn't work as expected.
- The "Segregation" Problem: In some cases, the computer predicted a perfect mix, but in reality, one ingredient (like Calcium) refused to mix in. It acted like a guest who wouldn't sit at the main table but instead stood in the corner with their own plate. The computer saw this tendency in the math, but the "perfect" mix still required temperatures so high that the salad would melt before it could form.
- The "Hidden Competitor" Problem: Sometimes, the ingredients preferred to pair up with each other to form a different type of dish entirely (like a ternary phase), rather than staying in the big five-way mix. The computer had to be updated to include these "cheaters" in its database to get the prediction right.

The Bottom Line

This paper proves that you can use a mix of physics simulations and AI to quickly screen thousands of potential new materials.

What it achieved: It created a ranking system to find the most promising candidates for new, stable, high-entropy oxides.
What it couldn't do: It couldn't give a perfect "cook at exactly 1,200°C" instruction because real-world chemistry is messy (magnetic effects, melting points, and hidden chemical reactions).
The Takeaway: Instead of blindly trying every combination, scientists can now use this digital tool to pick the top 10 or 20 most likely winners to test in the real lab, saving time and resources. It's like using a weather app to decide which day to have a picnic, rather than waiting outside for a week to see if it rains.

Problem Statement

Entropy-stabilized oxides (ESOs) offer access to vast multicomponent compositional spaces with potential for unprecedented properties. However, identifying promising candidates is challenging due to the sheer number of possible mixtures and the difficulty of assessing their thermodynamic stability against competing phases. Traditional brute-force experimental searches are time-consuming, and no general design rule exists for selecting favorable element sets. Specifically, the authors aim to screen equimolar quinary ( $n=5$ cations) ESOs in the NaCl (rocksalt) structure type, a system where mixing five ordered oxides does not always guarantee the stabilization of a single-phase ESO.

Methodology

The authors developed a high-throughput computational workflow integrating Density Functional Theory (DFT), Special Quasirandom Structures (SQS), convex-hull thermodynamics, and supervised machine learning.

Reference Database Construction:
- A consistent database of ordered binary and ternary oxides was constructed using 16 cation species (Ca, Cu, Co, Fe, Hf, In, Mg, Mn, Ni, Sn, Sr, Ti, V, Y, Zn, Zr).
- Calculations utilized both GGA (PBE) and meta-GGA (r2SCAN) functionals. Zero-point energy (ZPE) corrections were applied to ordered phases using Phonopy.
- The database includes various crystal prototypes (e.g., corundum, rutile, spinel) to ensure a robust ground-state convex hull.
Disorder Modeling (SQS):
- Atomic disorder in equimolar solid solutions was modeled using SQS generated via the Alloy Theoretic Automated Toolkit (ATAT).
- Cluster expansions included pair clusters up to the 7th nearest neighbor.
- Supercell sizes were optimized based on correlation root-mean-square error (RMSE) and DFT total energy convergence: 32 atoms for binaries, 48 for ternaries, and 60 atoms for quinary systems.
- ZPE for disordered mixtures was approximated as a linear combination of ordered end-member ZPEs, validated to introduce negligible error (<1%) in mixing enthalpy.
Thermodynamic Stability Assessment:
- Stability was assessed via a custom iterative convex-hull algorithm in a six-dimensional composition space ( $x = \{x_{M1}...x_{M5}, x_O\}$ ).
- The distance to the hull ( $\Delta_{hull}H$ ) determines stability at 0 K.
- Finite-temperature stabilization ( $T_{stab}$ ) was defined as the temperature where the Gibbs free energy of the ESO intersects the convex hull of competing phases, accounting for configurational entropy ( $\Delta S_{conf}$ ) approximated by Boltzmann's expression.
Machine Learning (ML) Screening:
- To screen all 4,368 possible equimolar quinary combinations ( $\binom{16}{5}$ ) affordably, supervised learning models were trained to predict $\Delta_{hull}H$ .
- The training set consisted of 416 DFT-calculated quinary SQS configurations (approx. 10% of the total space) alongside ordered unary, binary, and ternary data.
- Input features included cation fractions and physicochemical descriptors (atomic number, radius, electronegativity, etc.).
- Three models were benchmarked: Linear Regression, Random Forest, and Multilayer Perceptron (MLP).

Key Results

Database and Functional Performance: The meta-GGA (r2SCAN) functional provided significantly better agreement with experimental ground states and lower convex-hull energies (by 10–20 kJ/mol) compared to PBE, particularly for transition metal oxides.
SQS Validation: The chosen 60-atom supercells for quinary systems achieved correlation RMSEs near zero and energy fluctuations below 1 kJ/mol. The linear ZPE approximation was validated against full phonon calculations on a binary system.
Machine Learning Performance:
- The optimized MLP model achieved the best predictive performance with a test RMSE of 4.24 ± 0.19 kJ/mol.
- This model required explicit DFT calculations for only ~10% of the quinary systems to predict the properties of the full set of 4,368 compositions.
- The MLP outperformed Linear Regression (6.09 kJ/mol) and Random Forest (4.86 kJ/mol), highlighting the importance of non-linear correlations in multicomponent oxide stability.
Candidate Identification: The workflow identified 72 candidates with predicted $T_{stab} < 3,500$ K. The top 10 candidates (e.g., (Ca,Cu,Fe,Mg,Zn)O with $T_{stab} \approx 1,676$ K) were listed, showing reasonable chemical constraints (charge balance).
Experimental Validation:
- Experimental synthesis of selected candidates (e.g., (Cu,Fe,Mg,Sr,Zn)O and (Ca,Co,Cu,Mg,Zn)O) revealed that while some NaCl-type phases formed, they often exhibited cation segregation (e.g., Sr/Fe or Ca separating into distinct phases) or melted below the predicted $T_{stab}$ .
- The computational decomposition paths correctly predicted the tendency for elements with large size or chemical differences (like Ca or Sr) to segregate rather than form a single-phase quinary ESO.
- The model successfully recovered known ESOs (e.g., (Co,Cu,Mg,Ni,Zn)O) among the lowest predicted stabilization temperatures.

Significance and Claims

The paper claims to establish an efficient, data-driven route for exploring multicomponent oxides. Its primary contributions are:

Methodological Efficiency: Demonstrating that a supervised ML model, trained on a small subset of DFT data, can effectively screen thousands of complex quinary compositions, reducing the computational cost of high-throughput screening.
Physical Insight: The approach captures main stability trends and dominant competing phases, providing a physically meaningful ranking of new candidates. It confirms that configurational entropy alone is insufficient for stabilizing single-phase quinary NaCl ESOs; enthalpic factors (charge transfer, oxidation states) and competition with lower-order oxides are critical.
Practical Guidance: While absolute stabilization temperatures are affected by systematic thermodynamic approximations (e.g., neglect of magnetic entropy, liquid-phase thermodynamics), the framework provides practical guidance for experimentalists to prioritize compositions and rationalize synthesis outcomes, such as the need to avoid elements that cause premature melting or segregation.

The authors conclude that their workflow serves as an effective screening tool for prioritizing promising compositions, though they remain modest about the precision of absolute $T_{stab}$ values, noting they should be interpreted as comparative indicators rather than exact synthesis temperatures.