{"id":"c0d2f2b3-da75-4300-8c08-00927db5b584","arxiv_id":"1911.06836","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":1.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"A comprehensive review of multireference electron correlation methods and their analytical nuclear gradients for exploring potential energy surfaces.","lead":"This paper is a review of multireference quantum chemistry methods for potential energy surfaces, focusing on analytical nuclear gradients and derivative couplings. It is a reliable orientation for scientists deciding which multireference method and software to use for geometry optimizations and photodynamics simulations.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The practical-tool claim rests on the authors' own BAGEL FIC/XMS-CASPT2 gradients and couplings, which are asserted but not independently validated in this review; a systematic implementation error would propagate into the highlighted applications.","rationale":"The reader's weakest assumption is that the software implementations, especially the authors' FIC-CASPT2 and XMS-CASPT2 code in BAGEL, are correct and that the highlighted numerical comparisons are unbiased. I agree that this is the most load-bearing point. The review is a literature review, not a primary validation paper, so the absence of an independent numerical audit is a normal limitation rather than a fatal flaw. The central claim is additionally supported by many independent applications of MRCI and PIC-CASPT2 by other groups, by the open-source availability of BAGEL, and by the review's transparent discussion of approximations such as the use of CASSCF derivative couplings in some early MS-CASPT2 dynamics (Sec. III.B). The concrete check I propose would directly test whether the BAGEL FIC/XMS-CASPT2 gradients and couplings are correct; if they are, the review's conclusion stands. I therefore recommend no change to the reader's ACCEPT verdict, while noting that a validation supplement would strengthen the paper.","tokens_in":55525,"tokens_out":6983,"duration_ms":77469,"concrete_test":"Independently verify BAGEL's FIC-XMS-CASPT2 analytical gradients and derivative couplings on a small benchmark set (e.g., PSB3 or pyrrole) by (a) computing central-difference numerical gradients with the same method, active space, and basis set and requiring agreement to ~1e-6 a.u. for all 3N Cartesian components, and (b) cross-comparing with MOLPRO's PIC-(X)MS-CASPT2 analytical gradients at the same geometries. For derivative couplings, check the norm and the g.h overlap against an independent finite-difference or analytic implementation near a known conical intersection. Agreement at numerical precision would settle the concern; systematic deviations would require revising the strength of the practical-tool claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The review's central claim is that analytical nuclear gradients and derivative couplings make multireference methods practical for geometry optimization and on-the-fly photodynamics. The most distinctive recent pillar of that claim is the authors' own FIC-CASPT2 and XMS-CASPT2 analytical gradient and derivative coupling code in BAGEL (Refs. 17, 19, 63, 66), which underpins many of the 2016-2019 applications in Tables I and II (e.g., Refs. 63, 66, 68, 70, 71, 73, 74, 19, 20, 22). The manuscript describes the Lagrangian formalism and the automatic code-generation route (Secs. VI.D and VI.E) but does not include finite-difference validation, comparison against an independent implementation, or error analysis for the final gradient and coupling routines. If, for example, the lambda-equation solution or the CI-derivative tensor contractions in Eq. (101) contained a subtle bug, the reported conical intersection geometries, branching-plane topologies, excited-state lifetimes, and quantum yields could be systematically off, and the 'workhorse' conclusion would be overstated. This is a verification gap rather than a demonstrated error: BAGEL is open-source, the method papers are peer-reviewed, and several applications were run by independent groups. The review also notes that the PIC-MS-CASPT2 gradient by Celani and Werner has never been published (Sec. II.C), further limiting external verification of an important comparison baseline.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript is a review of multireference (MR) electron correlation methods for computing potential energy surfaces, with an emphasis on analytical nuclear gradients and derivative couplings. It first surveys geometry optimizations and on-the-fly nonadiabatic dynamics applications (Tables I and II), then introduces the electronic structure background (CASSCF, MRCI, MRCC, MRPT), and then presents the gradient and derivative-coupling formalism, including the Hartree-Fock/MCSCF foundations, the Lagrangian approach, and its application to PIC-CASPT2, FIC-CASPT2, and (X)MS-CASPT2. It closes with conical intersection optimization and semiclassical dynamics interfaces. The central claim is that the availability of analytical gradients and derivative couplings has made MR methods practical tools for geometry optimizations and photodynamics, even where single-reference methods fail.","tokens_in":55941,"tokens_out":12825,"duration_ms":139499,"significance":"The paper is a useful, broad reference that connects historical developments in MR gradient theory to modern applications. Its main strengths are the coherent exposition of the Celani-Werner Lagrangian, the coverage of internal-contraction variants and zeroth-order Hamiltonian choices, the documentation of the automatic code-generation route (smith3) for FIC-CASPT2, and the extensive tables of applications. The manuscript also credits open-source software (BAGEL) and cites many applications by groups other than the authors. The verification-gap concern of the stress-test does not, on reading the manuscript, amount to a demonstrated flaw; the cited method papers are peer-reviewed and BAGEL is open source, so the concern is a scoping caveat rather than evidence of incorrectness. Nevertheless, the most recent pillar of the 'practical tool' claim rests on the authors' own BAGEL implementations, and a brief, explicit statement about the absence of new independent benchmarking would make the claim more precise.","major_comments":[],"minor_comments":[{"comment":"The printed expression for the derivative coupling, h^{PT2,QP} = 1/2 [<Q|dΨ_P/dX> + <Ψ_Q|dP/dX>], is not the standard definition; for normalized real wave functions the plus-sign form would vanish. Please replace it with h^{QP} = <Ψ_Q|dΨ_P/dX> (or an explicitly antisymmetrized expression with a minus sign).","section":"§VI.E, Eq. (106)"},{"comment":"The stationarity condition is written as ∂L/∂U_rs - ∂L/∂U_rs = 0; the second term should be ∂L/∂U_sr.","section":"§VI.C, Eq. (79)"},{"comment":"The clause 'several reviews ... on the CI and CC theories' cites Refs. 6, 77, 140, and 175; Ref. 140 is a nonadiabatic dynamics review and does not belong in that citation grouping. Please correct the citation list.","section":"§IV.B"},{"comment":"Many entries in the System column are blank and rely on footnote letters to identify the molecules; this makes the otherwise valuable survey difficult to scan. Please fill the System column directly or provide an explicit list of systems in the caption.","section":"Tables I and II"},{"comment":"The manuscript explicitly notes that the PIC-MS-CASPT2 nuclear gradient 'has never been published' even though it underlies several early MS-CASPT2 dynamics applications. This is an acknowledged reproducibility limitation; the review should state its consequences or explicitly point to the later published implementations that supersede it.","section":"§II.C and §III.B"},{"comment":"The recent FIC-CASPT2 and XMS-CASPT2 applications cited as evidence of practicality depend on the authors' BAGEL implementation, but the review does not state that no independent numerical validation is performed here. A one-sentence caveat, with pointers to finite-difference checks in Refs. 17, 19, 63, and 66, would make the scope of the claim precise.","section":"§VI.D and §III.B"},{"comment":"The entries for the NEVPT2 gradient papers give only a year and no journal or preprint identifier; since Ref. 274 is a self-reference, please supply full bibliographic data or label both as preprints.","section":"References 273–274"},{"comment":"There are several typographical errors, including 'state-speficic' (§VI.F), 'graidents' (§V.B.2), 'earilest' (§III.A), 'unconvered' (§III.A), and 'some the studies' (§II.C). They should be corrected.","section":"General"}],"recommendation":"minor_revision","confidential_remarks":"The review is informative and the theoretical exposition is sound, but the self-citation concentration in the sections on FIC/XMS-CASPT2 is noticeable. A short transparency note about the use of the authors' own implementations and the absence of new independent benchmarks would be appropriate, in addition to the requested local corrections."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a solid, genuinely useful review. It won't reshape the field, but it will orient people who need to choose among multireference methods and understand how the gradients are actually computed. The main claim—that analytic gradients make these methods practical—is well supported by the applications literature.\n\nWhat's good: the theory sections are clear and historically accurate. The Lagrangian/Z-vector derivation is the right pedagogical backbone, and the coverage of internal contraction schemes (FIC vs PIC, XMS vs MS) is more explicit than most papers. I especially appreciate the candid note that the PIC-MS-CASPT2 gradient was never published; that kind of transparency is rare. The application tables are a useful quick reference, and the examples span organic photochemistry, nucleobases, transition metals, and actinides.\n\nWhere it's soft: the review leans heavily on the authors' own BAGEL FIC/XMS-CASPT2 work. That's not circularity in a damning sense—those papers are peer-reviewed and many applications come from independent groups—but a reader should know that the 'workhorse' label is partly a statement of faith in code that isn't independently benchmarked in this review. The stress-test note about a potential bug in the CI-derivative contractions is real as a verification gap, but it's not evidence of a bug, and it applies equally to any review that cites software results. I'd call that a minor limitation, not a load-bearing flaw.\n\nOther soft spots: minor typos ('graidents' in Sec. V), and the numerical results are taken on faith from the original papers. For a review, that's normal.\n\nWho this is for: graduate students and experimental collaborators who need to know when CASPT2 vs MRCISD vs XMS-CASPT2 is appropriate, and method developers who want a compact summary of gradient theory. It deserves a serious referee, and I'd accept it after minor revisions—mainly fixing typos and tempering a couple of 'workhorse' claims to acknowledge that the reliability rests on the underlying implementations.","headline":"A thorough, accurate review of multireference gradients and dynamics; the self-citation heavy narrative is a minor concern, not a flaw.","tokens_in":56283,"tokens_out":2252,"would_cite":true,"duration_ms":25692,"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":"Analytical gradients make multireference photodynamics practical.","keywords":["multireference electron correlation","analytical nuclear gradients","derivative couplings","CASPT2","conical intersections","on-the-fly photodynamics","potential energy surfaces","surface hopping"],"falsifier":"Take a small photochemical system with a known conical intersection, such as ethylene or a protonated Schiff base model, and optimize the S1/S0 crossing with both XMS-CASPT2 and a high-level uncontracted MRCI reference using the same basis set; a qualitative mismatch in the branching-plane vectors and seam topology would show that the practical reliability asserted by the review does not transfer to that system.","tokens_in":55355,"feed_emoji":"⚛️","tokens_out":5219,"duration_ms":51685,"temperature":0.7,"pith_summary":"This review argues that multireference electron correlation methods have matured from single-point energy tools into workhorses for exploring potential energy surfaces, because analytical nuclear gradients and derivative couplings are now available for them. The central message is that geometry optimization, including excited-state minima and conical intersections, and on-the-fly photodynamics simulations are practical at correlated levels such as CASPT2, even when single-reference methods fail. The authors trace the theory from Hartree-Fock gradients through the Lagrangian approach that makes multireference gradient derivations tractable, and they catalog applications showing where dynamical correlation changes the picture qualitatively.","feed_headline":"Analytical gradients make multireference photodynamics practical","feed_subtitle":"CASPT2 gradients and derivative couplings now drive geometry optimization and on-the-fly dynamics simulations.","key_machinery":"The central object is the analytical nuclear gradient expressed through the Lagrangian and the Z-vector equation. Because the CASPT2 or multistate CASPT2 energy is not stationary with respect to orbital rotations or CI coefficients, the gradient is obtained by constructing a Lagrangian with constraints, solving the Z-vector response equations, and forming relaxed (effective) densities. The derivative couplings follow from the interstate coupling term, which for multistate CASPT2 comes from differentiating the effective Hamiltonian, plus a determinant term. For fully internally contracted CASPT2, automatic code generation is the mechanism that made the involved CI derivatives tractable.","core_discovery":"The paper's central claim is that analytical nuclear gradients and derivative couplings for multireference electron correlation methods, above all for internally contracted CASPT2, have reached the point where they can be used routinely in geometry optimizations and nonadiabatic dynamics. The load-bearing development is the Lagrangian (response-function) formalism, which handles the fact that correlated multireference energies are not variational in the orbital and configuration-interaction coefficients by introducing Z-vector equations and effective densities; this is what makes the gradient evaluation practical. With fully internally contracted CASPT2, automatic code generation overcame the algebraic complexity of the CI derivative. The review documents that including dynamical correlation through these gradients changes conical intersection geometries, branching-plane topologies, and photodynamics lifetimes compared with CASSCF, often bringing calculations into agreement with experiment.","pith_inferences":["If analytical gradients stop being the bottleneck, the same Lagrangian-plus-automatic-derivation machinery could be carried over to other multireference methods such as NEVPT2 and multireference coupled cluster, and to higher-order properties like Hessians and spin-orbit couplings.","The systematic differences between CASSCF and CASPT2 conical intersections reported here suggest that CASSCF-based mechanistic conclusions in photochemistry should be rechecked with a correlated gradient method whenever the two states differ in dynamical correlation.","As implementations speed up, multireference on-the-fly dynamics may become a standard complement to time-resolved spectroscopy, turning simulation into an interpretive tool rather than a specialized calculation."],"forward_implications":["Geometry optimization of excited states and conical intersections can be carried out at the CASPT2 and XMS-CASPT2 levels, not only at CASSCF.","On-the-fly surface-hopping and ab initio multiple spawning simulations with multireference perturbation theory now reproduce experimental lifetimes and branching qualitatively, including intersystem crossing when spin-orbit coupling is added.","Choosing XMS-CASPT2 avoids artifacts from state rotations in multistate perturbation theory, giving reliable potential-energy-surface topologies near crossings.","MRCI and CASPT2 gradient data provide references against which cheaper methods such as TDDFT, ADC(2), and semiempirical MRCI can be benchmarked for conical intersection geometries.","Further progress depends on reducing cost; explicitly correlated and local-correlation multireference gradients are named as the next step."],"supporting_citations":[{"why":"Supplies the Lagrangian formulation and the first analytical CASPT2 gradient for partially internally contracted CASPT2.","marker":"[16]"},{"why":"Establishes automatic code generation for fully internally contracted CASPT2 nuclear gradients.","marker":"[17]"},{"why":"Extends the gradient formalism to XMS-CASPT2 with a state-rotation-invariant zeroth-order Hamiltonian.","marker":"[47]"},{"why":"Provides the multistate extension of fully internally contracted CASPT2 nuclear gradients.","marker":"[63]"},{"why":"Presents analytical derivative couplings for multistate CASPT2.","marker":"[66]"},{"why":"Demonstrates on-the-fly CASPT2 surface hopping with analytical gradients.","marker":"[19]"},{"why":"Shows XMS-CASPT2 dynamics for cyclohexadiene reproducing the experimental lifetime and quantum yield.","marker":"[22]"},{"why":"Provides the first ab initio multiple spawning dynamics with MS-CASPT2, establishing how derivative couplings were approximated there.","marker":"[108]"}],"fun_headline_variants":["CASPT2 gradients now drive routine photodynamics","Z-vector formalism enables practical CASPT2 gradients","Automatic code generation tames CASPT2 gradient algebra","Multireference gradients shift conical intersections vs CASSCF"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The review's conclusion that these methods are practical workhorses rests on the correctness of the software implementations and on the selected applications being representative; if the implementations contain bugs or the highlighted systems are atypical, the claimed practicality would be overstated.","fun_headline_variants_meta":{"raw":{"variants":["CASPT2 gradients now drive routine photodynamics","Z-vector formalism enables practical CASPT2 gradients","Automatic code generation tames CASPT2 gradient algebra","Multireference gradients shift conical intersections vs CASSCF"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000178,"raw_usage":{"total_tokens":1273,"prompt_tokens":896,"completion_tokens":377,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":512,"completion_tokens_details":{"reasoning_tokens":313}},"tokens_in":512,"tokens_out":377,"duration_ms":4471,"temperature":1.0,"reasoning_tokens":313,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:23:44.206729+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a small photochemical system with a known conical intersection, such as ethylene or a protonated Schiff base model, and optimize the S1/S0 crossing with both XMS-CASPT2 and a high-level uncontracted MRCI reference using the same basis set; a qualitative mismatch in the branching-plane vectors and seam topology would show that the practical reliability asserted by the review does not transfer to that system.","supporting_citations":[],"review_version":1}