{"id":"6b497af6-4144-4a04-8cae-2cd56eb19f98","arxiv_id":"2505.06431","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"SELFE computes magnetic ordering temperatures and heat capacity from DFT-derived exchange parameters combined with an efficient self-consistent single-site sampling scheme, matching established methods on BCC-Fe and FCC-Fe with up to 142 times fewer averaging steps.","lead":"This paper introduces SELFE, a computational method that predicts the magnetic heat capacity and magnetic ordering temperatures of materials using far fewer simulation steps than existing approaches. A materials scientist or engineer might read it because the method could make high-throughput screening for magnetic cooling and memory materials practical.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SELFE's accuracy claim rests on an unverified single-site sampling approximation and a validation loop that shares TB2J input; the FCC-Fe 4x overestimate leaves one experimental benchmark.","rationale":"The reader correctly identified the common TB2J input as a validation weakness. My stress-test adds a more internal concern: the self-consistent single-site sampler's stationary distribution is not proven to equal the Boltzmann distribution of the Heisenberg Hamiltonian in Eq. (4). The close SELFE/Multibinit agreement in Fig. 4 is evidence of internal consistency between two solvers, but it does not resolve whether the sampler introduces a systematic bias because both solvers consume identical exchange parameters. The FCC-Fe case reinforces this: both methods overestimate TN by roughly a factor of four, so the only unqualified experimental match is BCC-Fe. A direct comparison to classical Heisenberg Monte Carlo with identical Jij would isolate the sampler's correctness. If that comparison passes, the remaining limitation is one of validation breadth; if it fails, the central claim is unsupported. Since the reader's verdict is already CONDITIONAL and my concern strengthens rather than redirects that conclusion, I recommend no change to the verdict, with the Monte Carlo benchmark and a third-material test as explicit conditions.","tokens_in":14498,"tokens_out":8334,"duration_ms":93431,"concrete_test":"Run SELFE and a well-converged classical Heisenberg Monte Carlo (e.g., single-spin Metropolis with 10^5–10^6 sweeps after equilibration) on the same 5x5x5 BCC-Fe and FCC-Fe lattices using the same TB2J {Jij}, and compare cm(T), TC, and TN. If SELFE's transition temperatures and heat-capacity curves differ from Monte Carlo beyond statistical error, the single-site self-consistent sampler is not sampling Eq. (4) and the agreement with Multibinit is not evidence of accuracy. If the two agree, compute SELFE's TC/TN for a third material with a well-known experimental transition, e.g., Ni (TC ≈ 627 K), using the same automated workflow; experimental agreement would validate Jij transferability, while disagreement would localize the failure in the shared exchange parameters.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that SELFE is an accurate first-principles predictor rests on an unverified equivalence: the single-site distribution p(θ_i|β,h_i) in Eq. (2) is exact only for a Hamiltonian with fixed fields h_i, but SELFE applies it with h_i({e}) = |Σ_j J_ij e_j| and iterates to self-consistency. The stationary distribution of this self-consistent single-site sampler is not shown to equal the Boltzmann distribution of the Heisenberg Hamiltonian in Eq. (4). The only head-to-head validation compares SELFE to Multibinit on the same 5×5×5 lattice using the same TB2J {Jij}; Fig. 4 therefore tests two approximate solvers against each other, not the first-principles accuracy of the shared exchange parameters. For FCC-Fe both methods give TN ≈ 260 K against an experimental 67 K, leaving BCC-Fe as the sole experimental success. The reported 142x sample-efficiency gain is also measured against Multibinit's step counts, so the efficiency claim depends on the validity of the very comparison used as evidence. Thus neither the sampling approximation nor the transferability of TB2J Jij is independently established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Sampled Effective Local Field Estimation (SELFE), a method that combines DFT-derived Green's-function exchange parameters Jij with a self-consistent single-site angular sampling scheme to compute magnetic heat capacity and critical temperatures. The method is demonstrated on BCC-Fe and FCC-Fe, reporting TC = 1,016 K for BCC-Fe (experiment 1,043 K), TN = 259 K for FCC-Fe (experiment 67 K), close agreement with the Multibinit baseline (1,026 K and 256 K), and a claimed sample-efficiency improvement of up to 142x over Multibinit. The authors state that SELFE is fully automated, requires only a single unit-cell DFT calculation, avoids spin-constrained supercells, and has no free parameters.","tokens_in":14766,"tokens_out":5723,"duration_ms":60979,"significance":"If the central approximation is valid, SELFE would be a practically valuable method for high-throughput magnetothermal screening, since it avoids spin-constrained supercells and needs only one unit-cell DFT calculation. The paper's strengths include detailed DFT input files in the Supplemental Material, a clearly stated algorithm, and a convergence criterion that could be automated. However, the method's core statistical justification is not established, the experimental validation consists of one successful system, and the efficiency comparison is not yet well defined. Because the method is fast and reproducible, a revised version that addresses these points could be a useful contribution; in its current form the central claims outrun the evidence.","major_comments":[{"comment":"The key step of the method is the assertion that, 'Because the form is the same as the original one body Hamiltonian, we can use the same probability distribution above (Eqn. 2) to sample the magnetic moment angles.' This inference is not proven. Eq. (2) is exact for a Hamiltonian with independent fixed fields hi, but in SELFE the field is hi = Sum_j Jij e_j and depends on the current configuration of all spins; moreover, all sites are updated simultaneously from fields computed from the previous configuration. No derivation or numerical test shows that the stationary distribution of this self-consistent sampler equals the Boltzmann distribution of Eq. (4). This is a mean-field-like approximation, not an exact sampling scheme. The authors should either prove convergence to the correct equilibrium distribution or benchmark SELFE against an exact Monte Carlo / heat-bath sampling of Eq. (4) on the same lattice and with the same Jij. Without this, the first-principles accuracy claim for SELFE is not established.","section":"Section II, Eq. (4) and Algorithm 1"},{"comment":"The experimental validation currently rests on a single successful case. BCC-Fe gives TC = 1,016 K, close to the experimental 1,043 K, but FCC-Fe gives TN = 259 K versus the experimental 67 K, a factor of about four overestimate. The close agreement with Multibinit's 256 K is not independent validation, because both methods use the same TB2J exchange parameters computed from the same DFT electronic structure; the comparison tests two approximate solvers against each other, not the transferability of Jij to finite-temperature statistics. The abstract's claim of 'excellent agreement' with experimental data is therefore too strong. The authors should either expand the benchmark set to additional magnetic systems or substantially soften the accuracy claims in the abstract and conclusions.","section":"Section IV, Table I, Figure 4"},{"comment":"The claimed 'sample efficiency improvement of up to 142x' is not well defined. SELFE is reported to use 350 averaging steps and 1 thermalization step, while Multibinit uses 5,000 (BCC-Fe) or 2,000 (FCC-Fe) thermalization steps plus 50,000 or 10,000 averaging steps. However, SELFE's 350 steps appear to exclude the self-consistent iterations needed to reach the convergence criterion at each temperature, and the mixing parameter gamma and window size n also affect the total work. Without reporting the total number of Hamiltonian evaluations, including all SELFE self-consistency iterations, and preferably wall-clock times, the 'over two orders of magnitude' efficiency gain is not verifiable.","section":"Section III, Table I"},{"comment":"The Z-score convergence criterion uses overlapping energy differences: Delta_m = E_k - E_{k-m} for m = 1, ..., n, so the same E_k appears in multiple Delta_m values. These differences are not independent, which invalidates the statement that Z follows a Student's t-distribution with n-1 degrees of freedom. This could cause premature or delayed convergence detection. The authors should use non-overlapping blocks of energy differences or otherwise justify the independence assumption.","section":"Section II, Eqs. (5)-(7)"}],"minor_comments":[{"comment":"The text contains several typographical errors: 'Perdue-Burke-Ernzerhof' should be 'Perdew-Burke-Ernzerhof', 'Briollion Zone' should be 'Brillouin Zone', and 'wre' should be 'were'.","section":"Section III"},{"comment":"The heat capacity curves are normalized to [0,1], which removes the absolute magnitude of the magnetic specific heat. Since magnetothermal applications depend on the absolute value of cm, the authors should report unnormalized heat capacities or provide the normalization scale.","section":"Section IV, Figure 4"},{"comment":"The notation '1x1 g' and '1x2 g' for DFT requirements is ambiguous; it should be explicitly defined (e.g., one unit-cell calculation for BCC-Fe and one two-atom primitive calculation for FCC-Fe).","section":"Table I"},{"comment":"The paper states that SELFE 'eliminates the need for free parameters,' but Algorithm 1 requires user-specified values for the mixing parameter gamma, the confidence level alpha, the window size n, the maximum number of iterations nmax, and the spin-lattice size. These are algorithmic parameters rather than physical fitting parameters, but the claim should be phrased more precisely.","section":"Section II, Algorithm 1"},{"comment":"The derivation of the site-level distribution assumes the field is oriented along the z-axis in the local frame. The main text handles the general orientation with rotation matrices, but the supplemental derivation should state this assumption explicitly so the connection between Eq. (2) and the general vector case is clear.","section":"Supplemental Material, Eq. (16)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript presents a fast and potentially automatable method, but the central statistical approximation is unproven and the experimental benchmark set is very small, with only BCC-Fe being a success. I would like the authors to benchmark SELFE against an exact Monte Carlo simulation of Eq. (4) on the same lattice, report a complete efficiency accounting, and either expand the experimental validation or significantly soften the claims in the abstract. I do not see a fundamental flaw that would make the approach impossible to salvage, so major revision rather than rejection seems appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things. The paper is an honest method proposal that does something new, and the accuracy case is thinner than the abstract suggests. The BCC-Fe Curie temperature (1,016 K vs 1,043 K) is a genuine success; the FCC-Fe Néel temperature is 4x off experiment, and the paper says so.\n\nWhat's new: SELFE's combination of TB2J exchange parameters with a self-consistent single-site Boltzmann sampler, replacing spin-constrained supercells with one unit-cell DFT run plus 350 sampling steps. That combination is not in the cited papers. The Z-score convergence criterion is a sensible way to automate the loop, and the convergence study over 200 runs with different mixing is a nice robustness check. They also include DFT input files, which helps reproducibility.\n\nThe soft spots are real. Eq. (2) is exact only for fixed local fields; using it with fields recomputed from neighbors and iterating is not shown to reproduce the Boltzmann distribution of the Heisenberg model. No proof is offered, and no test against an exact reference. The head-to-head with Multibinit uses the same TB2J Jij and same 5×5×5 lattice, so the agreement in Fig. 4 is a common-input consistency check, not independent validation. That leaves one experimental benchmark (BCC-Fe) as support for \"first-principles accuracy.\" The 142x efficiency number is step-count, not wall-time, and no code is released, so the practical claim is hard to evaluate.\n\nNone of this sinks the paper. The idea is plausible and the implementation is tidy. But the abstract overclaims. A referee should ask for (1) a demonstration on at least one system where the exact Heisenberg result is known, or a direct test of the sampler's stationarity; (2) a material outside iron; and (3) released code or at least detailed pseudocode with convergence and timing. With those, it could be a genuinely useful tool for high-throughput magnetic screening.\n\nMy recommendation: send it to peer review. It deserves referee time, and it will come back stronger after revision. I wouldn't cite it yet as an established method, but I'd keep an eye on it.","headline":"SELFE is a promising efficiency trick for magnetic heat-capacity prediction, but its accuracy claim currently rests on one experimental match plus a shared-input comparison to Multibinit.","tokens_in":15268,"tokens_out":3223,"would_cite":false,"duration_ms":33331,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"SELFE computes magnetic transition temperatures from one DFT calculation and 350 samples, matching a spin-dynamics baseline at up to 142x lower sampling cost.","keywords":["Sampled Effective Local Field Estimation","magnetothermal properties","magnetic heat capacity","Curie temperature","Néel temperature","exchange interactions","Green's function method","self-consistent sampling"],"falsifier":"Run the method on a magnetic compound whose transition temperature is known from experiment and whose low-energy physics is dominated by nonlocal correlations, for example a frustrated triangular or pyrochlore antiferromagnet, and compare the predicted heat-capacity peak with experiment and with a long Monte Carlo simulation using the same $J_{ij}$. If the 350-sample SELFE curve misses the transition while the long simulation captures it, the single-site factorization in Eq. (2) is the point of failure; if SELFE still matches, the factorization is more robust than the derivation implies.","tokens_in":14319,"feed_emoji":"🧲","tokens_out":8414,"duration_ms":80712,"temperature":0.7,"pith_summary":"SELFE (Sampled Effective Local Field Estimation) is a way to compute temperature-dependent magnetic properties from a single DFT calculation instead of sampling hundreds of constrained spin supercells. The authors argue that if the exchange couplings between magnetic sites are known from a Green's function calculation, one can sample each spin's orientation from a one-site Boltzmann distribution under the field created by its neighbors, and iterate self-consistently. This replaces the free parameters of earlier single-site models with a first-principles input, making the workflow automatable. On BCC-Fe and FCC-Fe, the method matches the spin-dynamics package Multibinit's heat-capacity curves and transition temperatures while using only 350 averaging samples, an up to 142x reduction in sampling steps. If this holds for other magnetic materials, magnetic specific heat and ordering temperatures become cheap enough to screen in high-throughput materials discovery.","feed_headline":"SELFE cuts magnetic-material sampling from 50,000 steps to 350","feed_subtitle":"It reproduces iron's two magnetic transition temperatures using one DFT run and no free parameters.","key_machinery":"The load-bearing object is the single-site Boltzmann factor $p(\\theta_i \\mid \\beta, \\bar{h}_i) = \\frac{\\beta \\bar{h}_i}{4\\pi \\sinh(\\beta \\bar{h}_i)} e^{\\beta \\bar{h}_i \\cos \\theta_i}$, which gives the conditional distribution of the polar angle of each spin relative to its effective local field. SELFE sets that field by the Heisenberg-type sum $\\bar{h}_i = \\sum_j J_{ij} \\hat{e}_j$, with the isotropic $J_{ij}$ computed from a Green's function expression so that no empirical parameter enters. The algorithm iterates: sample angles from $p$, rotate into the global crystal frame, mix with the previous configuration, recompute $\\bar{h}_i$, and stop when a Z-score on recent energy differences passes a Student's $t$-distribution threshold. The factorization that lets the partition function become a product of independent single-site integrals is what makes 350 samples sufficient; it is also the mechanism that keeps the method fully automated.","core_discovery":"The paper's central claim is that a solid's magnetic transition temperature and magnetic heat capacity can be obtained from a single unit-cell DFT calculation, without spin-constrained supercells or user-set parameters, by sampling each spin from the conditional distribution $p(\\theta_i \\mid \\beta, h_i) = \\frac{\\beta h_i}{4\\pi \\sinh(\\beta h_i)} e^{\\beta h_i \\cos \\theta_i}$ under an effective local field $\\bar{h}_i = \\sum_j J_{ij} \\hat{e}_j$ built from Green's-function-derived exchange couplings. Applied to iron, the method gives $T_C = 1{,}016$ K for BCC-Fe, close to the experimental $1{,}043$ K, and $T_N = 259$ K for FCC-Fe, matching the Multibinit baseline's $256$ K while both exceed the experimental $67$ K estimate for that phase. The authors present the close agreement with Multibinit's heat-capacity curves and the 350-sample convergence as evidence that the single-site sampling under self-consistent local fields captures the finite-temperature magnetic statistics at a fraction of the usual cost.","pith_inferences":["The paper's comparison is between two implementations that consume the same exchange couplings, so the close match in Fig. 4 mainly shows that the two sampling strategies agree, not that either is independently validated against magnetism outside iron; a discriminating test would use a compound whose $J_{ij}$ were fixed by inelastic neutron scattering.","The reported 142x improvement counts averaging samples; it does not include the cost of the Green's function $J_{ij}$ calculation or the quadratic field updates, so a wall-clock comparison for large cells would be tighter.","The single-site form of Eq. (2) factorizes the spin statistics; on a frustrated lattice, where correlations cannot be captured by a local field, the method may underperform, and that is exactly where a stress test should be run."],"forward_implications":["High-throughput screening of magnetic materials becomes practical: one unit-cell DFT run per candidate supplies the $J_{ij}$, and 350 samples per temperature yield the ordering temperature and heat-capacity peak.","Both ferro- and antiferromagnetic ordering temperatures are covered by the same code path, so the method is not tuned to one magnetic order.","Because the effective field in Eq. (4) is the only material input besides the lattice, the accuracy ceiling for transition temperatures is set by the exchange couplings $J_{ij}$; any improvement in $J_{ij}$ transfers directly to SELFE.","The same sampling distribution can be reused for other thermal averages, so magnetic susceptibility, magnetocaloric entropy change, and magnetic contributions to the Gibbs free energy are reachable extensions of the same workflow."],"supporting_citations":[{"why":"Supplies the Green's function procedure used to compute the isotropic exchange couplings $J_{ij}$ from a single DFT ground state.","marker":"[26]"},{"why":"Implements that Green's function exchange-parameter calculation; produces the $J_{ij}$ fed to both SELFE and Multibinit.","marker":"[33]"},{"why":"Prior single-site sampling framework that SELFE extends; defines the distribution $p(\\theta)$ and is the baseline whose spin-constrained supercell cost SELFE removes.","marker":"[20]"},{"why":"Multibinit is the spin-dynamics baseline whose heat-capacity curves and critical temperatures SELFE matches with far fewer steps.","marker":"[29]"},{"why":"Provides the experimental reference for the BCC-Fe Curie temperature and a theoretical FFC-Fe Néel temperature used in comparison.","marker":"[35]"},{"why":"Provides the experimental estimate of 67 K for the FCC-Fe Néel temperature against which SELFE's overestimate is noted.","marker":"[36]"},{"why":"Introduces the single-site Hamiltonian form $H = -\\sum_i h_i \\cdot \\hat{e}_i$ whose partition function yields the sampling distribution in Eq. (2).","marker":"[23]"},{"why":"Shows a Heisenberg-type formulation with local quantization; cited for the form of the two-body Hamiltonian in Eq. (4).","marker":"[19]"}],"fun_headline_variants":["One DFT run, 350 steps: SELFE reproduces iron’s transitions","143x faster magnetic sampling, SELFE matches experiment","Single unit cell, no free parameters, magnetothermal curves: SELFE","SELFE: 350 spin samples instead of 50,000 for accurate transitions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the assumption that the exchange couplings $J_{ij}$ obtained from the Green's function procedure are accurate enough, and that treating each spin independently under its current local field captures the physics, so that the finite-temperature statistics of the real material are the statistics of Eq. (2).","fun_headline_variants_meta":{"raw":{"variants":["One DFT run, 350 steps: SELFE reproduces iron’s transitions","143x faster magnetic sampling, SELFE matches experiment","Single unit cell, no free parameters, magnetothermal curves: SELFE","SELFE: 350 spin samples instead of 50,000 for accurate transitions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000998,"raw_usage":{"total_tokens":4195,"prompt_tokens":883,"completion_tokens":3312,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":499,"completion_tokens_details":{"reasoning_tokens":3232}},"tokens_in":499,"tokens_out":3312,"duration_ms":22639,"temperature":1.0,"reasoning_tokens":3232,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:41:36.119523+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the method on a magnetic compound whose transition temperature is known from experiment and whose low-energy physics is dominated by nonlocal correlations, for example a frustrated triangular or pyrochlore antiferromagnet, and compare the predicted heat-capacity peak with experiment and with a long Monte Carlo simulation using the same $J_{ij}$. If the 350-sample SELFE curve misses the transition while the long simulation captures it, the single-site factorization in Eq. (2) is the point of failure; if SELFE still matches, the factorization is more robust than the derivation implies.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Green's function procedure used to compute the isotropic exchange couplings $J_{ij}$ from a single DFT ground state."},{"cited_title":"Ab initio calculation of the magnetic Gibbs free energy of materials using magnetically constrained supercells","cited_arxiv_id":"2202.11492","evidence_quote":"Prior single-site sampling framework that SELFE extends; defines the distribution $p(\\theta)$ and is the baseline whose spin-constrained supercell cost SELFE removes."},{"cited_title":"Spin polarized ground state DFT calculations were performed in Siesta [30]","cited_arxiv_id":null,"evidence_quote":"Multibinit is the spin-dynamics baseline whose heat-capacity curves and critical temperatures SELFE matches with far fewer steps."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the experimental reference for the BCC-Fe Curie temperature and a theoretical FFC-Fe Néel temperature used in comparison."},{"cited_title":"Hagl¨ of, A","cited_arxiv_id":null,"evidence_quote":"Provides the experimental estimate of 67 K for the FCC-Fe Néel temperature against which SELFE's overestimate is noted."},{"cited_title":"Gyorffy, A","cited_arxiv_id":null,"evidence_quote":"Introduces the single-site Hamiltonian form $H = -\\sum_i h_i \\cdot \\hat{e}_i$ whose partition function yields the sampling distribution in Eq. (2)."},{"cited_title":"Walsh, M","cited_arxiv_id":null,"evidence_quote":"Shows a Heisenberg-type formulation with local quantization; cited for the form of the two-body Hamiltonian in Eq. (4)."}],"review_version":1}