{"id":"f56c8a52-b3c0-4207-8892-1715e5df3ef4","arxiv_id":"2505.00845","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"For Fe_Al:AlN, DFT+cRPA and G0W0 embedding methods give excitation energies near the QMC reference but disagree with it on state symmetry and d-orbital occupation, due to DFT-inherited crystal-field splittings and screened interactions.","lead":"This paper uses quantum Monte Carlo to check cheaper embedding calculations for an iron defect in aluminum nitride, and finds that the embedding methods can match excitation energies while predicting the wrong d-electron wave functions. It matters because embedding methods are widely used to simulate defects for quantum technologies, and energy-only checks can hide errors that change the physical predictions.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"QMC reference not robust enough to anchor the central comparison: basis-set shifts and a ground-state-only Jastrow can reorder the states and d-orbital occupations used to diagnose embedding errors.","rationale":"The reader's weakest assumption identifies the QMC reference convergence as the load-bearing point, and the paper's own data support that concern. Table I shows a 0.37 eV basis-set shift in the 4A2 excitation energy between HSE06 vtz and vqz, which is the same size as the embedding-to-QMC energy differences that the paper interprets as meaningful. More importantly, Sec. VII A states that the fixed ground-state Jastrow changes the ordering of the lowest 4A1 and 4A2 states and lowers the 4E states by about 1 eV relative to CASCI, so the reference state ordering and orbital occupations are sensitive to a component of the trial wave function that is not optimized for the excited states. Since the paper's central diagnostic is not just excitation energies but the a1 orbital occupation in the excited states, an uncontrolled Jastrow bias could change the qualitative conclusion that embedding methods obtain the wrong orbital character. This concern is load-bearing but not disqualifying: the paper is transparent about the fixed Jastrow and basis-set choices, and the comparison is internally consistent, so the appropriate verdict remains CONDITIONAL. The recommended check, reoptimizing the Jastrow for each state, would directly settle whether the QMC reference is reliable enough to support the paper's central claim. The abstract's unfulfilled chromium claim is a separate presentation problem and does not change this assessment.","tokens_in":20117,"tokens_out":5246,"duration_ms":55477,"concrete_test":"Recompute the QMC reference at HSE06 vqz with Jastrow factors optimized separately for the 6A1, 4A2, and 4E states using the ensemble cost functional in Eq. (2), and then repeat the full analysis of Table IV. If the 4A2-4E ordering and the reported ⟨n_a1⟩ values change by more than about 0.1 eV or 0.1 electrons relative to the submitted values, the reference is not converged and the attribution of embedding errors to the one-body crystal-field splitting in Sec. IV C would need revision. If the ordering and occupations remain stable, the concern is resolved and the central conclusion stands.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV A selects HSE06 vqz orbitals for the QMC reference because they minimize the ensemble cost functional in Eq. (2), but Table I then shows the 4A2 excitation energy is 1.33(7) eV with HSE06 vtz and 1.70(9) eV with HSE06 vqz. The 0.37 eV shift is comparable to the \"few tenths of an eV\" agreement claimed for embedding methods in Sec. IV B, so the QMC reference does not currently resolve whether those methods' energies agree or disagree at the accuracy needed for the central claim. The sharper problem is the Jastrow factor. Section III B states that the same two-body Jastrow, optimized only for the ground state, is appended to every eigenstate; Sec. VII A reports that this fixed Jastrow lowers the lowest vertical excitations by roughly 1 eV and changes the ordering of the lowest 4A1 and 4A2 states relative to CASCI. Because the Jastrow is not state-specific, its bias need not cancel in energy differences, and it directly affects the a1 occupations in Table IV that are used to conclude that embedding models have \"different orbital character\" and that cRPA's interaction tensor is too non-spherical (Sec. IV C). The central inference therefore rests on an unquantified systematic error in the reference, not merely on the quoted single-sigma statistical errors.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper benchmarks quantum embedding methods (DFT+cRPA with different double-counting schemes and a G0W0-based embedding) against variational quantum Monte Carlo calculations for a substitutional Fe defect in AlN in a 32-atom supercell. The authors compute vertical excitation energies and Fe d-orbital occupations from both approaches, finding that several embedding models obtain excitation energies within a few tenths of an eV of the QMC reference but with different orbital characters and state orderings. They attribute the dominant error to the one-body crystal-field splitting inherited from DFT rather than to double-counting corrections, and further argue that cRPA overestimates the non-spherical part of the screened interaction. The QMC reference is built from multi-Slater-Jastrow trial wave functions optimized with an ensemble variational principle.","tokens_in":20422,"tokens_out":10131,"duration_ms":94629,"significance":"The paper makes a useful methodological contribution by comparing not just excitation energies but also one-particle density-matrix occupations and symmetry labels between embedding models and QMC for the same 32-atom Fe_Al:AlN cell. The use of the ensemble variational principle (Eq. (2)) to construct excited-state trial wave functions and the detailed tabulation of state data (Tables IV–IX) are strengths. If the QMC reference is reliable, the conclusion that energy agreement can be misleading and that the DFT one-body crystal field is a dominant error source is important for the defect-embedding community. However, because the reference itself carries a large unquantified systematic uncertainty, the quantitative and even qualitative conclusions are not yet firmly established.","major_comments":[{"comment":"The QMC reference 4A2 excitation energy shifts from 1.33(7) eV (HSE06 vtz) to 1.70(9) eV (HSE06 vqz), a variation of 0.37 eV that is comparable to the 'few tenths of an eV' agreement claimed between embedding and QMC in Sec. IV B. The selection of HSE06 vqz is based on minimizing the ensemble cost functional in Eq. (2), but this variational criterion does not control the systematic error of excitation energies, and the error bars shown in Fig. 3 include only single-sigma statistical noise. Consequently, the comparison in Fig. 3 cannot currently distinguish whether the no-DC DFT+cRPA 4E state at 1.35 eV agrees with the QMC 4E at 1.78(4) eV at the claimed accuracy, and the statement that embedding excitation energies agree 'within uncertainties' is not supported.","section":"Sec. IV A, Table I"},{"comment":"The same two-body Jastrow, optimized only for the ground state, is appended to every eigenstate in Eq. (1). Sec. VII A reports that this fixed Jastrow lowers the lowest vertical excitations by roughly 1 eV and changes the ordering of the lowest 4A1 and 4A2 states relative to CASCI. Because the Jastrow is not state-specific, its bias need not cancel in excitation energies, and it directly affects the a1 occupations in Table IV that are used in Sec. IV C to conclude that cRPA's interaction tensor is too non-spherical. The paper does not report how the a1 occupations of the 4E states change between CASCI and QMC, so the conclusion that embedding models have 'different orbital character' (Sec. IV B) is not shown to be robust to the Jastrow choice.","section":"Sec. III B and Sec. VII A"},{"comment":"The inference that QMC indicates a significantly more spherically symmetric Coulomb interaction than cRPA rests on the a1 occupation of 0.61 in the QMC 4E state (Table IV) and on the ligand-field model of Sec. VII D. That ligand-field model is fitted to the QMC splittings and occupations (B fixed at 0.068 eV and CFS parameters chosen to reproduce the QMC ordering), so it is not an independent confirmation. If the Jastrow bias changes the 4E a1 occupation from ~0.6 to ~1 (as the large Jastrow effect in Sec. VII A suggests), the inferred isotropy of the interaction would no longer hold. The authors should provide evidence that the orbital occupations in Table IV are stable under variation of the Jastrow form or basis set.","section":"Sec. IV C and Sec. VII D"},{"comment":"The abstract (both in the header and in the arXiv metadata) states that the paper assesses 'iron and chromium defects in aluminum nitride' and finds that 'the best double counting recipe is opposite in these two cases.' The full text, however, presents results only for Fe_Al:AlN and makes no mention of chromium. This is a missing-result claim that must be corrected; either the abstract must be revised to match the actual content or the chromium results must be added.","section":"Abstract and full text"}],"minor_comments":[{"comment":"The sentence 'the self-energy way be straightforwardly decomposed' in Sec. I contains a typo and should read 'may be straightforwardly decomposed'.","section":"Introduction"},{"comment":"The caption contains the misspelling 'uncertainities' and, more importantly, the phrase 'within uncertainties' should be qualified to indicate that only statistical uncertainties are meant, in light of the basis-set variation documented in Table I.","section":"Fig. 3 caption"},{"comment":"Table IV reports single-sigma statistical errors on excitation energies but not on the one-particle expectation values ⟨n_e⟩, ⟨n_e'⟩, ⟨n_a1⟩, and ⟨t_ee'⟩; since these expectation values are used to diagnose orbital character, error bars on them would help assess the significance of differences between methods.","section":"Table IV"},{"comment":"The paper describes the QMC results as 'variational best estimates' using the HSE06 vqz orbitals, but the ensemble cost functional in Eq. (2) is minimized over the choice of orbitals, not over the Jastrow or determinant coefficients; a brief statement of what was actually optimized would improve reproducibility.","section":"Sec. IV A"}],"recommendation":"major_revision","confidential_remarks":"The core idea of the paper is sound and the benchmark protocol is valuable, but the QMC reference has a large unquantified systematic uncertainty that is comparable to the effects being diagnosed. The authors should be asked to estimate this systematic error (e.g., through state-specific Jastrows, basis-set extrapolation, or a Jastrow-sensitivity analysis) and to revise the abstract so that it does not promise chromium results that are absent from the body. The current version overstates the reliability of the central comparison."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look before you commit referee time. The paper does something genuinely useful: it benchmarks embedding methods for Fe_Al:AlN against QMC and, crucially, compares d-orbital occupations, not just excitation energies. The finding that several embedding methods land within a few tenths of an eV of QMC but with clearly different wavefunction character is a real contribution, and the evidence that the one-body crystal-field splitting inherited from DFT is the dominant error source is well supported by the one-particle energies in Table II and the cRPA interaction anisotropy in Table III. The ligand-field analysis in Sec. VII D is a nice, honest interpretive tool.\n\nThe main soft spot is the QMC reference itself. The 4A2 excitation shifts from 1.33(7) eV (HSE06 vtz) to 1.70(9) eV (HSE06 vqz) — a 0.37 eV spread comparable to the \"few tenths of an eV\" agreement the paper highlights. The fixed Jastrow, optimized only for the ground state, lowers vertical excitations by roughly 1 eV and changes the ordering of the 4A1/4A2 states relative to CASCI. That is a known systematic bias, and the paper does report it, but it is not quantified in a way that lets the reader judge whether the a1 occupations (the basis of the orbital-character conclusion) are stable. The stress-test note is right that this bias could in principle affect those occupations; the authors should either run state-specific Jastrows or at least show that the occupations are insensitive to the Jastrow.\n\nSecond, the abstract promises iron and chromium defects and a two-case double-counting story that the body does not deliver — the full text contains iron only. That needs fixing before publication.\n\nThe 32-atom cell is small, and the authors acknowledge finite-size effects; the 72-atom cRPA result actually strengthens their argument about orbital delocalization, so I don't see that as a fatal flaw.\n\nBottom line: the central qualitative conclusion — energy agreement can hide character mismatch — is likely robust, but the quantitative QMC anchor is shakier than the paper's tone suggests. It deserves peer review, with requests for Jastrow sensitivity checks and an abstract correction.\n\nI'd cite this if I worked on defect embedding; the occupation diagnostic is a useful template. For a reading group, it's a maybe — specialized, but the methodological lesson is general.","headline":"Useful benchmark with a genuine new occupation diagnostic, but the QMC reference carries unquantified systematic error and the abstract overclaims a chromium case.","tokens_in":20985,"tokens_out":3033,"would_cite":true,"duration_ms":34562,"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":"For a strongly correlated iron defect in aluminum nitride, standard quantum embedding methods reproduce QMC excitation energies to within a few tenths of an electron-volt while getting the wave-function character of the lowest excited…","keywords":["quantum embedding","quantum Monte Carlo","point defects","strongly correlated electrons","double counting","crystal field splitting","excited states","Fe in AlN"],"falsifier":"Run the QMC reference with a complete-basis extrapolation and a Jastrow factor re-optimized for each state instead of fixed from the ground state; if the lowest 4E state then has $\\langle n_{a_1}\\rangle \\approx 1$ (doubly occupied $a_1$) or the 4A2/4E ordering flips, the claimed orbital-character mismatch collapses. Alternatively, if an embedding variant that changes only the double-counting term (leaving DFT orbitals unchanged) corrects both energies and occupations, the attribution to the crystal-field splitting is refuted.","tokens_in":19911,"feed_emoji":"⚛️","tokens_out":9687,"duration_ms":80662,"temperature":0.7,"pith_summary":"Quantum embedding methods are cheaper than full many-body treatments of strongly correlated point defects, but they involve uncontrolled approximations. This paper benchmarks several standard embedding variants against quantum Monte Carlo (QMC) for a neutral iron impurity in aluminum nitride, using the same 32-atom cell and the same active space of five iron d orbitals. The central finding is that several embedding methods reproduce the QMC excitation energies to within a few tenths of an electron-volt while predicting the wrong wave-function character: the lowest excited states have the wrong symmetry ordering and a doubly occupied $d_{z^2}$ orbital, whereas QMC finds that orbital less than singly occupied. The paper identifies the dominant error as the one-body crystal-field splitting inherited from the DFT starting point, with double-counting corrections and the cRPA-screened interactions playing secondary roles. The practical message is that matching excitation energies is not enough; embedding methods must also be validated against wave-function properties.","feed_headline":"Matching defect energies can hide wrong wave functions","feed_subtitle":"For Fe in AlN, the error lives in the DFT crystal-field splitting, not the double-counting correction.","key_machinery":"The load-bearing objects are the five iron d-like orbitals that form the active space, classified under the $C_{3v}$ point group as an $e$ pair, an $e'$ pair, and an $a_1$ orbital, and the occupation of the $a_1$ orbital in each many-body state, which serves as a compact wave-function-character diagnostic. The QMC reference is built from multi-Slater-Jastrow trial wave functions (a Jastrow factor multiplying a sum of Slater determinants) with a Jastrow factor fixed from the ground state, optimized with the ensemble variational principle, a weighted sum of state energies plus an overlap penalty. The embedding models are single-impurity Hamiltonians with one-particle hoppings from DFT or $G_0W_0$ and two-particle interactions from cRPA, with three choices of double-counting correction. The comparison runs on identical Hamiltonians and active spaces, so disagreements must come from the embedding approximations rather than model differences. The mechanism of the argument is the systematic comparison of excitation energies and $a_1$ occupations between QMC and each embedding variant, backed by ligand-field-theory analysis showing which occupations a spherically symmetric interaction would produce.","core_discovery":"The paper's central claim is that for the neutral iron substitutional defect in AlN ($\\mathrm{Fe}_{\\mathrm{Al}}^0$:AlN), several standard quantum embedding schemes—DFT plus constrained random phase approximation (cRPA) with and without Hartree double counting, and a $G_0W_0$-based embedding—produce low-lying excited states whose excitation energies are within a few tenths of an eV of the QMC reference but whose orbital character is qualitatively different. The QMC ground state is a high-spin $^6A_1$ state and the lowest excited states are $^4A_2$ and $^4E$ with $a_1$ occupation around 0.5–1.0, whereas the embedding models place $^4E$ states with a nearly doubly occupied $a_1$ orbital lowest, or in the Hartree-DC case produce an incorrect $^2A_1$ ground state. The paper argues that the dominant error is the one-body crystal-field splitting in the iron d orbitals inherited from the DFT starting point, not the double-counting correction, and that the cRPA-screened interaction is itself distorted by over-delocalized DFT orbitals. The abstract further states that the best double-counting recipe is opposite for iron and chromium defects, so no universal correction exists.","pith_inferences":["We infer that if the crystal-field-splitting error is truly dominant, embedding methods that build one-body terms from a more localized or self-consistent starting point should improve both energies and occupations; this is testable beyond the paper's claims.","The same diagnostic—comparing a symmetry-resolved orbital occupation such as $\\langle n_{a_1}\\rangle$—could, we suggest, be applied to other transition-metal and rare-earth defects where DFT orbital delocalization varies across the active space.","We further infer that benchmarks ignoring dynamical correlation (e.g., CASCI-only comparisons) may systematically misattribute embedding errors, since the paper shows the fixed Jastrow factor shifts excitation energies by roughly 1 eV and changes state ordering.","A concrete extension we propose is repeating the QMC-vs-embedding comparison in the 72-atom cell mentioned in the paper, where cRPA interactions become more spherical; the paper reports the embedding side but not a QMC reference there."],"forward_implications":["Excitation energies are not a sufficient benchmark for embedded defect models; two methods can agree on energies while disagreeing on the many-body wave function, so comparisons must include observables such as orbital occupations.","Efforts to improve embedding for strongly correlated defects should focus on correcting the one-body crystal-field splitting inherited from DFT—for example through better starting points—rather than on refining double-counting corrections.","cRPA screening computed from DFT orbitals does not automatically fix the orbital-delocalization error; the screened interaction tensor inherits the non-spherical distortion of the DFT orbitals.","The optimal double-counting correction is system-dependent (opposite for Fe and Cr in this family), so a universal DC prescription is unlikely to work for transition-metal defects.","QMC calculations on small supercells can serve as a practical reference standard for identifying which embedding approximations are reliable."],"supporting_citations":[{"why":"Supplies the DFT+cRPA embedding method for defects and the earlier finding that Hartree double counting gives a wrong low-spin ground state for Fe_Al:AlN.","marker":"[25]"},{"why":"Provides the G0W0-based embedding method and its double-counting correction used as one of the benchmarked schemes.","marker":"[37]"},{"why":"Supplies the ensemble variational principle used to optimize the QMC excited-state trial wave functions.","marker":"[50]"},{"why":"Provides the ligand-field-theory multiplet structure used to interpret state orderings and a1 occupations.","marker":"[51]"},{"why":"Gives the prior G0W0 embedding result for Fe_Al:AlN with the correct ground state and a lattice relaxation energy used to estimate the experimental vertical excitation energy.","marker":"[10]"},{"why":"Reports the experimental zero-phonon-line energy used to sanity-check the QMC vertical excitation estimate.","marker":"[70]"},{"why":"Demonstrates in vanadocene that active-space, static-screening, and double-counting choices each shift excitation energies by at least 0.5 eV, motivating this benchmark.","marker":"[39]"}],"fun_headline_variants":["Energies match, wave functions don't: embedding's hidden flaw","DFT crystal-field splitting, not double counting, is the embedding error","Matching defect energies can still mean wrong ground states","No universal double counting: Fe and Cr need opposite recipes","For Fe in AlN, embedding errors stem from DFT orbitals, not interactions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole comparison rests on the QMC calculation being accurate enough to fix the true state ordering and d-orbital occupations, but the paper's own data show the 4A2 excitation moving from 1.33(7) eV to 1.70(9) eV when the basis set changes and the fixed Jastrow factor altering the state ordering relative to CASCI, so the reference carries an unquantified systematic bias.","fun_headline_variants_meta":{"raw":{"variants":["Energies match, wave functions don't: embedding's hidden flaw","DFT crystal-field splitting, not double counting, is the embedding error","Matching defect energies can still mean wrong ground states","No universal double counting: Fe and Cr need opposite recipes","For Fe in AlN, embedding errors stem from DFT orbitals, not interactions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000337,"raw_usage":{"total_tokens":1897,"prompt_tokens":1013,"completion_tokens":884,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":629,"completion_tokens_details":{"reasoning_tokens":795}},"tokens_in":629,"tokens_out":884,"duration_ms":8096,"temperature":1.0,"reasoning_tokens":795,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:33:53.456626+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the QMC reference with a complete-basis extrapolation and a Jastrow factor re-optimized for each state instead of fixed from the ground state; if the lowest 4E state then has $\\langle n_{a_1}\\rangle \\approx 1$ (doubly occupied $a_1$) or the 4A2/4E ordering flips, the claimed orbital-character mismatch collapses. Alternatively, if an embedding variant that changes only the double-counting term (leaving DFT orbitals unchanged) corrects both energies and occupations, the attribution to the crystal-field splitting is refuted.","supporting_citations":[{"cited_title":"Muechler, D","cited_arxiv_id":null,"evidence_quote":"Supplies the DFT+cRPA embedding method for defects and the earlier finding that Hartree double counting gives a wrong low-spin ground state for Fe_Al:AlN."},{"cited_title":"Sheng, C","cited_arxiv_id":null,"evidence_quote":"Provides the G0W0-based embedding method and its double-counting correction used as one of the benchmarked schemes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the ensemble variational principle used to optimize the QMC excited-state trial wave functions."},{"cited_title":"Sugano, Y","cited_arxiv_id":null,"evidence_quote":"Provides the ligand-field-theory multiplet structure used to interpret state orderings and a1 occupations."},{"cited_title":"Strongly correlated states of transition metal spin defects: the case of an iron impurity in aluminum nitride","cited_arxiv_id":"2501.16280","evidence_quote":"Gives the prior G0W0 embedding result for Fe_Al:AlN with the correct ground state and a lattice relaxation energy used to estimate the experimental vertical excitation energy."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Reports the experimental zero-phonon-line energy used to sanity-check the QMC vertical excitation estimate."},{"cited_title":"Chang, E","cited_arxiv_id":null,"evidence_quote":"Demonstrates in vanadocene that active-space, static-screening, and double-counting choices each shift excitation energies by at least 0.5 eV, motivating this benchmark."}],"review_version":1}