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

Celine Barreteau; David Berardan; Jean-Claude Crivello; Sebastien Junier; Yann-Andrev Kerneur

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

T0 review · 3 major / 0 minor · reviewed 2026-07-11 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A neural network trained on ~10% of DFT data ranks all 4368 equimolar quinary NaCl-type entropy-stabilized oxides by stabilization temperature and recovers the known ones.

desk verdict Solid, usable NaCl-ESO screening pipeline with honest limits; MLP ranking is useful for prioritization but absolute Tstab and order can shift when the ordered hull is incomplete. read the letter →

arxiv 2607.04502 v2 pith:EXUIL76D submitted 2026-07-05 cond-mat.mtrl-sci cond-mat.dis-nn

Sebastien Junier , Celine Barreteau , David Berardan , Yann-Andrev Kerneur , Jean-Claude Crivello This is my paper · ORCID

classification cond-mat.mtrl-scicond-mat.dis-nn

keywords high-throughputcomputationalworkflowsupervisedlearninghigh-entropyoxidesentropy-stabilizedoxiderocksaltstructurespecialquasirandomstructuresconvexhullmultilayerperceptron

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper builds a high-throughput workflow that combines density-functional calculations, special quasirandom structures, convex-hull thermodynamics, and supervised learning to screen every equimolar five-cation NaCl-type oxide that can be formed from a library of 16 metals. After DFT on only about 416 of the 4368 compositions, an optimized multilayer perceptron predicts the distance of each mixture to the convex hull (and therefore its ideal stabilization temperature) with a test error of roughly 4 kJ per mole. The resulting ranking places the experimentally known entropy-stabilized oxides among the lowest predicted temperatures and supplies a short list of new candidates. Experimental synthesis trials and computed decomposition paths confirm that the model correctly identifies the dominant competing phases, even though absolute temperatures remain systematically high because of thermodynamic approximations. The practical payoff is a ranked map that experimental groups can use to decide which multicomponent oxides are worth attempting first.

What carries the argument

The iterative finite-temperature convex hull constructed from a consistent DFT database of ordered binary/ternary oxides plus SQS-modeled disordered quinary cells, with an MLP that maps composition plus simple chemical descriptors directly onto ΔhullH (and therefore Tstab).

What would settle it

Synthesize several of the lowest-ranked predicted candidates under conditions that avoid premature melting; if none form a single-phase rocksalt solid solution while several higher-ranked compositions do, or if the observed secondary phases systematically disagree with the computed decomposition paths, the ranking claim fails.

Watch

Extended reading notes

Core claim

An optimized multilayer perceptron, trained on special-quasirandom-structure DFT formation enthalpies for only about 10 percent of the 4368 equimolar quinary NaCl-type oxides, predicts the distance to the convex hull with a test RMSE of approximately 4 kJ mol^{-1} and thereby ranks all compositions by ideal stabilization temperature; the ranking recovers known entropy-stabilized oxides among the lowest temperatures and correctly anticipates the secondary phases observed in synthesis trials.

Load-bearing premise

Stabilization temperature is defined solely by ideal Boltzmann configurational entropy on the cation sublattice while magnetic, vibrational and liquid-phase contributions are ignored and the ordered reference database is treated as complete enough to fix the hull.

Editorial extensions

If this is right

Experimental groups can prioritize the short list of lowest-Tstab equimolar compositions rather than sampling the full 4368-space at random.
Known NaCl-type entropy-stabilized oxides reappear near the top of the predicted ranking, giving a concrete check that the workflow is not inventing spurious candidates.
Decomposition-path calculations supply the identity of the competing phases that must be suppressed, guiding non-equimolar or off-stoichiometric adjustments.
The same trained model can be queried for any new five-cation subset drawn from the 16-element library without additional DFT, lowering the cost of exploring nearby composition space.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Because absolute Tstab values are systematically high, the practical value of the ranking may lie more in relative order than in the numerical temperatures themselves; re-ranking after adding missing stable ternaries or liquid free energies would be a direct next test.
The same SQS-plus-MLP pattern could be transferred to other structure types (spinel, fluorite, perovskite) once an analogous ordered reference set is built, potentially generalizing the screening strategy beyond rocksalt.
Departing from strict equimolarity for oversized cations such as Ca, as the experimental case studies already hint, may be the fastest route from the ranked list to genuine single-phase samples.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

This is a careful high-throughput screen of equimolar quinary NaCl-type ESOs, not a new physical principle. What is new is the consistent end-to-end package: binary/ternary ordered hulls (PBE + r2SCAN + ZPE), 60-atom SQS for ~416 quinaries, an MLP that hits ~4.2 kJ/mol test RMSE on held-out quinaries from composition + simple descriptors, iterative Tstab over all 4368 compositions, and a few experimental case studies that check competing phases.

They do the hard parts properly. Same DFT settings across the reference set, SQS size and pair-cluster choices are justified by energy and correlation convergence, ZPE mixing is checked on (Mg,Sr)O, and the MLP is cross-validated with quinaries held out. Known ESOs land among the low-Tstab candidates, and the three experimental decompositions (Sr+Fe, Ca-containing systems) match the computed secondary phases once the missing SrFeO3−x is added. Code for the hull is promised; that helps.

The soft spots are real but the authors mostly flag them. Absolute Tstab is systematically high under GGA (thousands of K) and still imperfect under r2SCAN; ideal cation-only Boltzmann entropy and missing magnetic/vibrational terms are part of that. More important for ranking: the hull is only as complete as the ordered binary/ternary database. Adding SrFeO3−x reordered facets and raised Tstab for whole families by thousands of kelvin. Other unenumerated ternaries or liquids can do the same, so the order of the 72 candidates below 3500 K is not guaranteed to be stable. The ~4 kJ/mol MLP error is secondary to that post-processing vulnerability. Melting before equimolar single-phase formation also appears in the experiments; they note it.

This is for people who actually synthesize or screen high-entropy oxides and want a ranked shortlist plus a transparent workflow, not for someone looking for a foundational entropy theory. Math and data look solid for what they claim (relative ranking under a fixed hull). I would send it to peer review; the limits are stated clearly enough that referees can push on hull completeness and absolute-temperature language without killing the paper. Worth engaging if you work in this space.

Referee Report

3 major / 0 minor

Summary. The manuscript presents a high-throughput computational workflow that combines DFT (GGA and selective meta-GGA), special quasirandom structures (SQS), custom convex-hull thermodynamics, and supervised learning to screen all 4368 equimolar quinary NaCl-type entropy-stabilized oxides formed from 16 cations. A consistent ordered binary/ternary reference database is built, 416 quinary SQS cells are computed, and an optimized multilayer perceptron is trained to predict the distance to the convex hull (test RMSE ≈ 4.24 kJ mol⁻¹). These predictions are converted into stabilization temperatures Tstab via an iterative free-energy hull that uses ideal cation-only configurational entropy. Known ESOs appear among the lowest-Tstab candidates, and three experimental case studies show that computed decomposition products largely match observed secondary phases, although absolute Tstab values remain systematically high.

Significance. If the relative ranking of candidates is robust, the work supplies a practical, data-efficient route for prioritizing experimental synthesis of NaCl-type ESOs and a reusable reference database plus open convex-hull code. The explicit recovery of known ESOs, the quantified ML error on held-out quinaries, and the direct comparison of decomposition paths with synthesis outcomes are concrete strengths that go beyond pure high-throughput enumeration. The approach is therefore of clear interest to the high-entropy-materials community even if absolute temperatures remain approximate.

major comments (3)

§2.3 Eqs. (4–7) and §3.5.2: Tstab is obtained by feeding MLP-predicted ΔhullH into an iterative hull that uses ideal Boltzmann entropy on the cation sublattice only and a finite ordered binary/ternary reference set. The SrFeO3−x example shows that adding one previously omitted ternary raises Tstab of all Sr+Fe compositions by thousands of kelvin and reorders the hull facets. Because the same incompleteness can exist for other unenumerated ternaries (or liquid phases), the relative ordering of the 72 candidates with Tstab < 3500 K (Table 4 / Supp. F) is not demonstrated to be stable. A sensitivity analysis—re-ranking after systematic addition of known competing ternaries or after a controlled incompleteness test—is needed before the list can be presented as reliable synthesis guidance.
§3.2 and §3.5: Absolute Tstab values remain far from experiment even after meta-GGA correction (e.g., GGA 2745 K → meta-GGA 1249 K versus experimental ~1100 K for (Co,Cu,Mg,Ni,Zn)O; still higher for other systems). The paper correctly attributes this to ideal entropy, neglected magnetic/vibrational contributions, and missing liquid phases, yet still ranks candidates by these absolute numbers. Either a calibrated relative metric (e.g., ΔTstab relative to a known ESO benchmark under identical approximations) should be adopted for ranking, or the manuscript must quantify how large an error in ΔhullH or ΔSconf is required to invert the order of the top candidates.
§2.2 and binary hulls (Supp. B): Magnetic ordering for Co/Fe/Mn/Ni oxides is initialized ferromagnetically and not optimized; the text notes residual uncertainties and that r2SCAN only recovers key phases when experimental ordering is enforced. Because many low-Tstab candidates contain these elements, the effect of magnetic ground-state choice on both the reference hull and the SQS energies should be quantified for at least a representative subset, or the ranking should be restricted to non-magnetic compositions until this uncertainty is bounded.

Circularity Check

0 steps flagged · score 1.0 of 10

No load-bearing circularity: MLP predicts independent DFT ΔhullH labels; Tstab ranking is a thermodynamic post-process checked against external synthesis and known ESOs.

full rationale

The derivation chain is self-contained and non-circular. Formation enthalpies of ordered binaries/ternaries and of 416 quinary SQS supercells are obtained from first-principles DFT (PBE/r2SCAN) under a uniform protocol (Eq. 1, §2.2). Supervised models (LR/RF/MLP) are trained to map composition + chemical descriptors onto those independent DFT ΔhullH labels; 4-fold CV holds out ~104 quinaries never seen in training, yielding a genuine test RMSE of 4.24 ± 0.19 kJ/mol (Table 3, Fig. 4). The MLP is then applied to the remaining ~90 % of the 4368 compositions. Tstab is obtained by feeding the predicted (or DFT) ΔhullH into an iterative convex-hull solver that uses the ideal cation-only Boltzmann entropy (Eqs. 4–7); this is a thermodynamic definition, not a fit to experimental Tstab. Known ESOs (CoCuMgNiZn)O and (CoFeMgMnNi)O reappear among the lowest predicted Tstab, and three experimental case studies recover the dominant secondary phases predicted by the hull (SrFeO3−x, CaO segregation, etc.). The only self-references are ordinary methodological reuse (ATAT SQS, prior oxide databases, authors’ GitHub hull code) and do not force any numerical result. Absolute Tstab values are acknowledged to be approximate because of missing ternaries, neglected magnetic/vibrational entropy and liquid phases (§3.5.2, §4), but that is a completeness/approximation issue, not circularity: the ranking is not defined by construction from the quantities it claims to predict. Score 1 reflects only the trivial presence of self-citations that are not load-bearing.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The central ranking rests on standard DFT/SQS thermodynamics plus ideal mixing entropy and a finite ordered-phase library. Hyperparameters of the MLP are free parameters of the surrogate, not of the physics. No new particles or forces are invented; the main fragility is completeness of the hull database and the ideal-entropy Tstab definition.

free parameters (3)

MLP architecture and regularization (two hidden layers 60/80, tanh, SGD, L2=24.45)
Chosen by 4-fold CV on the quinary set; they control the surrogate error that is then used for all 4,368 rankings.
SQS supercell sizes (32/48/60 atoms for binary/ternary/quinary)
Selected by RMSE and energy convergence on benchmarks; residual SQS error (~1–2 kJ/mol) enters every DFT label.
Pair-cluster cutoff (first seven neighbor shells; no triplets)
Fixed after binary tests; defines the disorder model for all mixes.

assumptions (6)

domain assumption Configurational entropy of equimolar quinary NaCl ESOs is ideal Boltzmann on the cation sublattice only, with oxygen ordered, and magnetic/vibrational entropy can be neglected for ranking Tstab.
Eqs. (4)–(7) and §2.3 define Tstab from this ΔSconf; absolute temperatures are known to be high vs experiment.
domain assumption A finite library of ordered binary and ternary oxides (plus selected disordered binaries) is complete enough to place the 0 K convex hull for relative ESO screening.
§2.3, §3.1; §3.5.2 shows missing SrFeO3−x strongly revised Tstab for Sr–Fe compositions.
domain assumption SQS supercells with the chosen pair correlations represent the random solid solution well enough that residual energy error is ~1 kJ/mol.
§2.1, §3.2 and Fig. 1b justify sizes used for training labels.
domain assumption ZPE of disordered mixes is a linear combination of ordered end-member ZPEs (negligible mixing ZPE).
Eq. (2); validated only on (Mg,Sr)O binary (§2.2).
domain assumption GGA (PBE) rankings of ΔhullH are sufficiently correlated with meta-GGA and experiment for prioritization, even if absolute energies differ by 10–20 kJ/mol.
Large-scale screening and ML training use PBE; r2SCAN used selectively for validation (§2.2, §3.4–3.5).
standard math Standard DFT total-energy differences and convex-hull geometry correctly order phase stability at 0 K within the chosen functional.
Eqs. (1)–(3); conventional materials thermodynamics.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Data-Driven Prediction of NaCl-Type Entropy-Stabilized Oxide Compositions from First-Principles and Supervised Learning." pith.science (2026). https://pith.science/paper/EXUIL76D

@misc{pith2026260704502,
  author       = {Pith},
  title        = {Pith review of: Data-Driven Prediction of NaCl-Type Entropy-Stabilized Oxide Compositions from First-Principles and Supervised Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EXUIL76D}},
  note         = {Machine review of arXiv:2607.04502}
}

read the original abstract

Entropy-stabilized oxides (ESOs) open access to vast multicomponent compositional spaces, but identifying promising candidates remains challenging because of the large number of possible mixtures and the need to assess their stability against competing phases. In this work, we develop a high-throughput computational framework to screen equimolar quinary ESOs in the NaCl structure type by combining density functional theory (DFT), special quasirandom structures (SQS), convex-hull thermodynamics, and supervised machine learning. A consistent reference database of binary and ternary ordered oxides, including disordered phases such as all binary cation combinations in the NaCl-type oxide, is first constructed using GGA and meta-GGA calculations. Quinary disordered phases are then described by SQS supercells and used to train machine-learning models that predict the distance to the convex hull and the corresponding stabilization temperature over the full set of 4368 possible equimolar quinary compositions generated from 16 cation species. Among the tested models, an optimized multilayer perceptron provides the best predictive performance, with a test error of about 4 kJ/mol, while requiring explicit DFT calculations for only about 10% of the quinary systems. Comparison with experimental synthesis tests and computed decomposition paths further shows that the approach captures the main stability trends and the dominant competing phases, although absolute stabilization temperatures remain affected by systematic thermodynamic approximations. These results establish an efficient route for the data-driven exploration of multicomponent oxides and provide practical guidance for the experimental search for new ESOs.