{"id":"f5cd284d-9c3b-49b8-a5de-68aa6a93e5ab","arxiv_id":"2608.07445","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A permutation-invariant neural network predicts strain- and chemistry-dependent vacancy formation energies in FCC CoCrFeNi with R^2 = 0.958, showing volumetric strain dominates the average response.","lead":"This paper trains a machine learning model on half a million atomistic simulations to predict how stretching, compressing, or shearing a metal alloy changes the energy needed to create an empty atomic site. It finds that volume change is the main driver of the average energy shift, while the local mix of atoms still controls how much sites differ from one another.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Omitted removed-species identity is the key risk: the model may be predicting species-averaged rather than site-specific formation energies, and no per-species diagnostic is provided.","rationale":"Reading in good faith, the central physical observation that volumetric strain dominates the average vacancy-formation-energy response, with shear of minor influence, is supported by the atomistic calculations and the power-law fits in Table 1, and it does not depend on the machine-learning model. The neural-network claim is also plausible on its face: 500,000 EAM samples, a held-out R^2 of 0.958, a validation MAE of 0.026 eV, and an independent-configuration plot all provide real support. The important weak point is the species-blind representation. I partially agree with the reader: the mathematical indistinguishability argument assumes that the features are constructed from neighbor identities and positions that do not encode the removed species. If the local features are taken from the relaxed defective supercell, the removed species may be imprinted in the neighbor displacements, so the failure is not logically guaranteed. However, the paper does not demonstrate that this imprint is sufficient, and the design choice after Eq. (28) makes the question central. Because species-dependent chemical potentials enter Eq. (2), a model that cannot identify the removed species cannot achieve the claimed per-site accuracy unless the geometry happens to be a reliable species fingerprint. The proposed ablation test would settle whether the omission is benign or whether the model is only predicting a species average. The verdict should remain conditional rather than accept or reject: the physical trend is credible, but the species-resolved machine-learning claim needs explicit validation.","tokens_in":15608,"tokens_out":9387,"duration_ms":114390,"concrete_test":"Retrain the same architecture with the removed species z_i appended as a one-hot or learned site feature, holding the data split, strain sampling, and hyperparameters fixed. Compare per-species held-out MAE and R^2 against the published species-blind model. If the species-informed model lowers per-species MAE by substantially more than the published global MAE of 0.026 eV, the published representation is averaging over removed species rather than resolving them, and the claimed site-specific accuracy is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The training target in Eq. (2) is species-dependent through the chemical potential mu_A(F), but the per-neighbor feature vector in Eq. (10) is s_ij = [z_j, r_ij, a_ij]; no z_i or equivalent site-species embedding is present. The authors state explicitly after Eq. (28) that the removed species is not an independent input and is instead inferred from the surrounding environment. In a random solid solution without short-range order, the removed atom's identity is, to order 1/N, independent of the neighbor species counts, so a function of neighbor species alone cannot reproduce the species-dependent part of the target. If the input neighborhood is taken from the relaxed defective configuration, the removed species can in principle leave a trace in r_ij and a_ij through relaxation, but the paper gives no evidence that this trace is strong enough or unique; per-species errors are not reported. Aggregate R^2 = 0.958 and MAE = 0.026 eV do not establish species-resolved accuracy, and the subsequent site-specific enthalpy and vacancy-concentration analyses inherit this limitation. This is the load-bearing premise for the claim that the network accurately predicts vacancy formation energies across diverse deformation states.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a permutation-invariant Deep Sets surrogate for site-resolved vacancy formation energies in FCC CoCrFeNi under uniaxial and shear deformation, using vacancy-centered neighbor descriptors and invariants of the deformation gradient. The authors generate roughly 500,000 EAM vacancy formation energies per strain state, fit an empirical power law for the species-averaged volumetric-strain dependence, and report aggregate validation metrics of R² = 0.958 and MAE = 0.026 eV. The surrogate is then used to compute pressure-dependent formation enthalpies, site occupation probabilities, configurational entropies, and equilibrium vacancy concentrations, with the central claim that volumetric deformation dominates the average response while shear deformation is minor and that local chemical environments control the statistical distribution.","tokens_in":15902,"tokens_out":5840,"duration_ms":64939,"significance":"If the reported accuracy survives an independent-configuration test and is verified per removed species, the surrogate would be a practically useful fast evaluator for stress-dependent vacancy energetics in HEAs. The paper contributes a large EAM dataset, a clean rotation- and permutation-invariant local representation with a proof of rotational invariance, and a direct comparison against a second random solid-solution configuration. The main caveats are that the current quantitative claims are aggregate and that the thermodynamics of the entropy/concentration analysis is not well posed.","major_comments":[{"comment":"The training target in Eq. (2) contains the species-dependent chemical potential µ_A(F), but the node feature in Eq. (10) is s_ij = [z_j, r_ij, a_ij], which encodes the neighbor species and geometry but not the identity of the removed atom. In a random solid solution without short-range order, the removed species is statistically independent of the neighbor-species counts, so two sites with identical neighbor environments but different removed species receive identical predictions while the true formation energies differ by the species chemical potential. Relaxation of the defective configuration could in principle leave a trace of the removed species in r_ij and a_ij, but the manuscript provides no per-species error analysis to show that this trace is strong or unique. The aggregate R² and MAE in §4.2 therefore do not establish the claimed species-resolved accuracy; per-species parity plots or the explicit addition of the vacancy-site species as an input are needed.","section":"§2.2, Eq. (10) and remark after Eq. (28)"},{"comment":"The reported validation metrics come from a random 70/15/15 split of samples within the same supercell configurations and repeated strain states. Neighboring vacancy sites within a supercell share overlapping environments, and the same deformed cell contributes many samples, so the test set is not statistically independent. The only independent check, the 'new configuration' curves in Figure 4(b), is presented visually without any quantitative error measure. The authors should report MAE, RMSE, and R² computed on the fully unseen supercell configuration, ideally per strain state and per species, to support the transferability claim.","section":"§4.2 and Figure 4"},{"comment":"Eq. (41) defines a global Shannon entropy of the site-occupation probabilities p_i, not a site-specific vacancy formation entropy S_i^vf. Inserting this quantity into Eq. (39) conflates the equilibrium occupation distribution with the formation entropy of a single vacancy and double-counts the Boltzmann factor already present in p_i through Eq. (40). As a result, the absolute vacancy concentrations in Figure 7 are not grounded in a well-defined formation entropy. The authors should either derive a proper site-specific formation entropy or explicitly relabel Figures 6(b) and 7 as relative propensities rather than quantitative vacancy concentrations and formation entropies.","section":"§2.4, Eqs. (39)-(41)"}],"minor_comments":[{"comment":"The captions for Figure 6 are incorrect: panel (a) is labeled as 'true versus predicted values of vacancy formation energy' and panel (b) as 'predicted values versus volumetric strain,' while the actual panels show vacancy occupation probability and configurational entropy versus pressure.","section":"Figure 6"},{"comment":"It is not specified which checkpoint produced the results in Figure 4: the minimum validation loss occurs at epoch 908, the minimum validation MAE at epoch 978, and the minimum RMSE at epoch 972; these are three different models.","section":"§4.2"},{"comment":"The denominator N_j in the angular descriptor is ambiguous: it should be the number of partners k contributing to the sum, typically |N_i|-1 when j is excluded, rather than an undefined neighbor count for atom j.","section":"Eq. (12)"},{"comment":"The power-law fit in Eq. (8) is stated to describe the data well, but the table reports only the fitted constants and no goodness-of-fit measure such as per-species RMSE; adding this would strengthen the claim.","section":"Table 1"},{"comment":"There is no data or code availability statement. Given the size of the dataset and the central role of the neural-network implementation, providing trained weights and the data-generation pipeline would be important for reproducibility.","section":"General"},{"comment":"There are several small typographical issues, including 'site-speccific' in §2.4 and the capitalization of 'Where' after Eq. (39); these should be corrected during revision.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The decisive technical issue is the omitted removed-species identity. If the authors can provide per-species validation on fully unseen configurations and demonstrate that the model does not merely reproduce species-averaged behavior, the paper's central claim would be credible. The entropy formulation in Eqs. (39)-(41) also needs substantial reworking. I would be willing to review a revised version that addresses these points."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth a look for the physics observation and the architecture idea, but the central accuracy claim is not yet supported. The volumetric-dominance result comes from direct EAM calculations, not the ML model, and the power-law fit in Eq. 8 is a reasonable compact descriptor for species-averaged behavior. The combination of Cauchy-Green invariants with a vacancy-centered Deep Sets model for HEA defect energetics is genuinely new as far as I know, and the rotational invariance proof is straightforward and correct. The scale of the dataset (500k per strain state) is serious, and the validation numbers (R² = 0.958, MAE = 0.026 eV) are not trivial.\n\nThe soft spot is not a nitpick. Equation (2) defines the target with a species-dependent chemical potential μ_A(F), but the feature vector in Eq. (10) contains only neighboring species, distances, and angles. The identity of the removed atom is not an input, and the authors explicitly say so after Eq. (28). In a random equiatomic alloy, neighboring species counts are essentially independent of the removed species, so two sites with identical neighborhoods but different removed atoms will get identical predictions while their true formation energies differ by the chemical potential difference. The theoretical escape is that the relaxed geometry around the vacancy could in principle encode the removed species, but the paper gives no per-species error analysis to show the network actually uses that trace. Without it, the aggregate R² could hide a systematic failure to separate Fe vs Cr vs Ni vs Co vacancies. The downstream site-specific enthalpy, entropy, and concentration analysis in Figures 5–7 inherits this limitation.\n\nOther issues are smaller: the sample-level 70/15/15 split within supercells means the test set contains near-duplicate environments, so the generalization claim is weaker than the number suggests; the new-config check is graphical only; and there is no code or data to reproduce the training. The configurational entropy substitution (Shannon entropy of occupation probabilities in place of the true formation entropy) is acknowledged as an approximation, fine for an enthalpy-based estimate.\n\nWho gets value: anyone building strain-dependent defect surrogates for multiscale modeling, and anyone thinking about what descriptors are necessary for site-resolved predictions in chemically complex alloys. The paper deserves a serious referee. I would ask the authors to add the removed-species identity as a feature (or justify its absence with per-species CEs and error bars), report per-species MAE, and release the trained model and data. That is a substantial but well-scoped revision, not a desk reject.","headline":"A useful new descriptor combination for strain-dependent vacancy energetics in HEAs, but the omitted removed-atom species identity is a real gap that undermines the site-specific accuracy claim.","tokens_in":16397,"tokens_out":2928,"would_cite":false,"duration_ms":30499,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"In an FCC high-entropy alloy, vacancy formation energy shifts with volume, not shear, and a permutation-invariant network resolves the site-to-site chemistry.","keywords":["high-entropy alloys","vacancy formation energy","permutation invariant neural network","finite deformation","volumetric strain","machine learning surrogate","CoCrFeNi","defect energetics"],"falsifier":"Take two supercells in which the first three coordination shells around the vacancy site are identical but the removed central atoms are different species, compute the EAM vacancy formation energies directly, and compare them with the network's predictions; if the reference energies differ by the species chemical potential while the predictions are identical, the central premise fails.","tokens_in":15397,"feed_emoji":"⚛️","tokens_out":9912,"duration_ms":95725,"temperature":0.7,"pith_summary":"The paper shows that in the face-centered cubic high-entropy alloy CoCrFeNi, the average change in vacancy formation energy under mechanical load is controlled by volumetric strain, while shear strain contributes little. It further claims that the large site-to-site scatter at a fixed load is a chemical-environment effect, and that a permutation-invariant neural network can predict this site-resolved value. If true, this gives a fast surrogate for the expensive atomistic calculations needed in multiscale models of diffusion, irradiation damage, and shock failure. The support is a network trained on 500,000 EAM vacancy energies across 17 uniaxial and 17 shear strain states, with held-out $R^2=0.958$ and mean absolute error $0.026$ eV.","feed_headline":"Volumetric strain, not shear, sets vacancy formation cost","feed_subtitle":"Neural network matches atomistic vacancy energies to 0.026 eV and shows local chemistry controls the spread.","key_machinery":"The load-bearing object is the vacancy-centered descriptor $\\mathbf{s}_{ij}=[z_j, r_{ij}, \\mathbf{a}_{ij}]$, where $z_j$ is the neighbor species embedding, $r_{ij}$ is the distance from the vacancy, and $\\mathbf{a}_{ij}$ contains three averaged angle features (mean angle, mean cosine, mean $\\cos 2\\theta$) over neighbors around the vacancy. Summing the encoded neighbor contributions $\\xi_i=\\sum_j \\mathbf{m}_{ij}$ gives permutation invariance, and the three principal invariants $I_1, I_2, I_3$ of the right Cauchy–Green tensor $\\mathbf{C}=\\mathbf{F}^T\\mathbf{F}$ form an objective deformation descriptor; these are concatenated into $\\eta_i=[\\xi_i, I_1, I_2, I_3]$ and passed through dense layers. Because the removed atom's species is not an explicit input, the network must learn it from the surrounding shell. Equation (8), $E_{vf}=a(1+\\varepsilon_{\\mathrm{vol}})^{-n}+c$, is the simple analytic mechanism by which the paper explains the volume-dominated average behavior.","core_discovery":"The central discovery is that vacancy formation energy in FCC CoCrFeNi separates into a species-averaged, volume-controlled response and a chemistry-controlled site scatter. Under uniaxial loading the mean follows $E_{vf}=a(1+\\varepsilon_{\\mathrm{vol}})^{-n}+c$ with species-dependent constants, while simple shear leaves the average essentially unchanged; under identical macroscopic loading, individual sites still differ because each vacancy sits in a distinct chemical neighborhood. The network uses a vacancy-centered representation — neighbor species embeddings, radial distances, averaged bond-angle features, and the principal invariants of $\\mathbf{C}=\\mathbf{F}^T\\mathbf{F}$ — summed over neighbors for permutation invariance. Training on 500,000 EAM vacancy energies across 17 uniaxial and 17 shear states yields validation $R^2=0.9578$ and MAE $0.026$ eV, and the same volume trend appears for an independently generated configuration. The authors conclude that hydrostatic pressure shifts formation enthalpy and vacancy concentration without erasing the chemical hierarchy between sites.","pith_inferences":["Inference: because the representation contains no alloy-specific constants, the same architecture should transfer to other FCC high-entropy alloys, but the omitted-species assumption will be the first thing to test when species chemical potentials differ strongly.","Inference: making the removed species an explicit input and adding a species-potential correction would extend the model to ordered or short-range-ordered alloys, where the neighbor shell no longer determines the central atom.","Inference: the volume-only pressure dependence implies that shock or irradiation models could approximate vacancy-mediated damage from the material's volumetric compression history alone, using the fitted power-law coefficients."],"forward_implications":["A species-averaged power law $E_{vf}=a(1+\\varepsilon_{\\mathrm{vol}})^{-n}+c$ with fitted constants can replace expensive defect calculations when only the mean vacancy formation energy under load is needed.","Because shear has little effect, continuum and shock models can treat vacancy formation energy as a function of volume or pressure rather than the full strain tensor.","Site-resolved predictions from the network make it possible to write vacancy diffusion and irradiation-damage models with a distribution of formation energies instead of a single Arrhenius value.","At 300 K the model predicts equilibrium vacancy concentrations that change by many orders of magnitude between strong tension and strong compression, with local chemistry controlling the offset."],"supporting_citations":[{"why":"Provides the embedded-atom method interatomic potential used to generate every vacancy formation energy in the training set.","marker":"[55]"},{"why":"Introduces the Deep Sets permutation-invariant architecture that the network's summation pooling is built on.","marker":"[45]"},{"why":"Supplies the atomistic simulation engine used for energy minimization and cell relaxation under applied deformation.","marker":"[54]"},{"why":"Generates the random solid-solution atomic configurations used as the chemical environments.","marker":"[53]"},{"why":"Established that macroscopic deformation modifies vacancy formation energy, the effect this paper extends to high-entropy alloys.","marker":"[33]"},{"why":"Shows from first principles that local chemical environment governs vacancy energetics in CoCrFeMnNi, motivating the representation.","marker":"[23]"},{"why":"Provides an experimental vacancy formation enthalpy in a related high-entropy alloy against which the computed range is compared.","marker":"[22]"}],"fun_headline_variants":["Volume strain dominates vacancy cost, shear barely matters","Neural net predicts vacancy energy under strain and chemistry","Local chemistry controls vacancy scatter under same loading","Fast AI matches atomistic vacancy energies, MAE 0.026 eV","Vacancy formation: volume sets mean, chemistry sets spread"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The model assumes the removed atom's species is redundant because the surrounding neighbor shell determines the formation energy; if two identical neighborhoods with different removed species have different true energies, the network cannot represent the difference.","fun_headline_variants_meta":{"raw":{"variants":["Volume strain dominates vacancy cost, shear barely matters","Neural net predicts vacancy energy under strain and chemistry","Local chemistry controls vacancy scatter under same loading","Fast AI matches atomistic vacancy energies, MAE 0.026 eV","Vacancy formation: volume sets mean, chemistry sets spread"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000211,"raw_usage":{"total_tokens":1426,"prompt_tokens":972,"completion_tokens":454,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":588,"completion_tokens_details":{"reasoning_tokens":375}},"tokens_in":588,"tokens_out":454,"duration_ms":4829,"temperature":1.0,"reasoning_tokens":375,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T04:26:35.067752+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take two supercells in which the first three coordination shells around the vacancy site are identical but the removed central atoms are different species, compute the EAM vacancy formation energies directly, and compare them with the network's predictions; if the reference energies differ by the species chemical potential while the predictions are identical, the central premise fails.","supporting_citations":[{"cited_title":"Model interatomic potentials and lattice strain in a high-entropy alloy","cited_arxiv_id":null,"evidence_quote":"Provides the embedded-atom method interatomic potential used to generate every vacancy formation energy in the training set."},{"cited_title":"Order parameter engineering for random sys- tems.High Entropy Alloys & Materials, pages 1–14, 2023","cited_arxiv_id":null,"evidence_quote":"Generates the random solid-solution atomic configurations used as the chemical environments."},{"cited_title":"Electronic structure study regarding the influence of macro- scopic deformations on the vacancy formation energy in aluminum.Mechanics Research Communications, 99:58–63, 2019","cited_arxiv_id":null,"evidence_quote":"Established that macroscopic deformation modifies vacancy formation energy, the effect this paper extends to high-entropy alloys."},{"cited_title":"Defect energetics for diffusion in crmnfeconi high- entropy alloy from first-principles calculations.Computational Materials Science, 170:109163, 2019","cited_arxiv_id":null,"evidence_quote":"Shows from first principles that local chemical environment governs vacancy energetics in CoCrFeMnNi, motivating the representation."},{"cited_title":"Vacancy formation enthalpy in cocrfemnni high-entropy alloy.Scripta Materialia, 176:32–35, 2020","cited_arxiv_id":null,"evidence_quote":"Provides an experimental vacancy formation enthalpy in a related high-entropy alloy against which the computed range is compared."}],"review_version":1}