{"id":"3bcd1a50-a8be-4d0f-b8f5-11b05bfe178e","arxiv_id":"2506.03811","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"DD-vMCG quantum dynamics on trans-PSB3 predicts full trans-cis twist at about 250-300 fs, delayed compared to AIMS and surface-hopping benchmarks, with CH stretches required for isomerisation in reduced models.","lead":"This paper simulates a small retinal-like molecule, PSB3, with a fully quantum wavepacket method and finds that its light-driven twist from trans to cis takes roughly 250-300 femtoseconds, much slower than earlier trajectory-based simulations reported. The slowdown is attributed to the wavepacket being trapped near an excited-state minimum before it reaches the conical intersection that returns the molecule to the ground state.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline comparison is not method-only: the cited benchmarks used different electronic structure methods, so the delayed isomerisation may reflect the SA(2)-CAS(6,6)/6-31G surface rather than DD-vMCG dynamics.","rationale":"The reader's weakest-assumption analysis identifies the same load-bearing issue: the central method comparison is contaminated by differences in electronic structure treatment. This is the single most important point because the paper's headline claim is explicitly a method comparison ('DD-vMCG shows a trans-cis isomerisation hundreds of femtoseconds slower using the same electronic structure method'). If that premise fails, the claim reduces to a single-surface simulation with an unexplained delay, not a method-level finding. The paper does have genuine strengths: the full-dimensional DD-vMCG simulation is a significant computational effort, the data are deposited, and the reduced-dimensionality runs showing CH-stretch dependence are internally informative. Those strengths do not, however, rescue the uncontrolled benchmark comparison. The proposed concrete test is the minimal controlled experiment that would decide whether the delay is due to the nuclear dynamics method or to the SA(2)-CAS(6,6)/6-31G surface. Because the reader's CONDITIONAL verdict already reflects exactly this uncertainty, no change to the verdict is needed.","tokens_in":11568,"tokens_out":3526,"duration_ms":37218,"concrete_test":"Run an independent trajectory-based simulation (e.g., Tully fewest-switches surface hopping or AIMS) on the identical SA(2)-CAS(6,6)/6-31G surface with all 36 degrees of freedom, using Wigner sampling from the same ground-state normal modes and the same initial S1 vertical excitation. Compare the time-dependent S1 population and the average C2-C3-C4-C5 torsion against Figure 3. If the trajectory method also shows S1 decay only after ~200 fs and cis formation at 250-300 fs, the delay is a property of the PES, not of DD-vMCG; if it decays within ~60 fs as in the cited benchmarks, the delay is attributable to the quantum wavepacket method.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that DD-vMCG shows trans-cis isomerisation 'hundreds of femtoseconds slower using the same electronic structure method' than trajectory methods, with the delay attributed to wavepacket effects and accessibility to the conical intersection. But the comparison is confounded. Section 2.1 states that the Martínez AIMS work used a three-state average, Szymczak et al. used SA(2)-CASSCF followed by MRCI, and the present work uses plain SA(2)-CAS(6,6)/6-31G without post-CASSCF correlation and without the S2 state. The abstract's 'same electronic structure method' is therefore not accurate, and the two candidate explanations for the delay—(i) the nuclear dynamics method and (ii) the chosen PES—are not separated. The barrier invoked in Figure 5 is shown on arbitrary-unit axes and is itself generated under the current SA(2)-CASSCF surface, so it cannot independently establish that the delay is method-induced. The residual 60% S1 population at 300 fs further complicates the interpretation of the 'fully isomerised' final state. The load-bearing assumption is that the SA(2)-CAS(6,6)/6-31G surface is close enough to the benchmark surfaces that the timescale difference is caused by DD-vMCG; this assumption is untested and is exactly what the central claim depends on.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript reports on-the-fly direct dynamics variational multiconfigurational Gaussian (DD-vMCG) simulations of the trans-PSB3 retinal model in full dimensionality (36 normal modes), using a two-state-averaged CAS(6,6)SCF/6-31G electronic structure method. The central result is that the S1-to-S0 population transfer occurs only after about 200 fs and that the average torsional angle reaches cis-like values by 250-300 fs, which is hundreds of femtoseconds slower than previous AIMS and surface-hopping benchmarks. The authors attribute this delay to trapping in the S1 trans minimum and to a barrier en route to the conical intersection, and they use reduced-dimensionality runs to argue that CH stretch modes are essential for successful isomerisation.","tokens_in":11805,"tokens_out":4899,"duration_ms":47156,"significance":"If the central claim holds, this is a useful contribution to the method-comparison literature for nonadiabatic dynamics: it is, to my knowledge, the first full-dimensional on-the-fly vMCG study of PSB3, and the delayed timescale is an output of the propagation rather than a fitted parameter. The paper also ships a large electronic-structure database and Zenodo-hosted data, and it explicitly compares its population and geometry time series with well-known benchmarks, which is valuable for reproducibility. The main weakness is that the headline comparison is not method-only: the benchmark trajectory calculations used different electronic structure treatments, so the delayed isomerisation may be a property of the SA(2)-CASSCF surface rather than of quantum wavepacket dynamics. This confound must be resolved or carefully qualified before the paper's central attribution can be accepted.","major_comments":[{"comment":"The central claim that DD-vMCG shows isomerisation 'hundreds of femtoseconds slower using the same electronic structure method' is not supported by the text. Section 2.1 states that the Martínez AIMS benchmark used a three-state average and that Szymczak et al. used SA(2)-CASSCF followed by MRCI, while the present work uses plain SA(2)-CAS(6,6)/6-31G without post-CASSCF correlation. The delayed timescale could therefore reflect the different potential energy surfaces (missing S2 state, missing dynamic correlation) rather than the nuclear dynamics method. Please either qualify the claim, run a trajectory method on the identical SA(2)-CAS(6,6)/6-31G surface, or provide quantitative evidence that the relevant surface features (S1 minimum, barrier height, CI accessibility) are comparable across these electronic-structure choices. The barrier in Figure 5 is computed on the present PES and cannot independently resolve this confound.","section":"Abstract; Section 2.1"},{"comment":"The conclusion that PSB3 'has fully isomerised' by 250-300 fs is in tension with Figure 3(a), which shows that about 60% of the population remains in S1 at 300 fs. The average torsional angle and BLA in Figures 3(c) and 3(d) are weighted averages over Gaussian centers and can mask a bimodal wavepacket (for example, trans-like population on S1 and cis-like population on S0). The reported standard deviations describe the spread of Gaussian centers, not a state-resolved nuclear density. Please report state-resolved torsional and BLA distributions, or at least diabatic-population-weighted histograms, and rephrase the 'fully isomerised' claim to reflect that the average structure becomes cis-like while substantial S1 population remains.","section":"Section 3.1; Conclusions"},{"comment":"The database histogram is used to infer that the wavepacket is 'trapped' in the S1_trans minimum and that two deactivation routes exist near MECI_db and MECI_opt. However, the density of stored electronic-structure points is not a probability distribution of the nuclear wavepacket; it reflects on-the-fly sampling, database interpolation queries, and Hessian-update decisions. The observation that 3.6% of the 48,857 structures lie near S1_trans does not by itself establish trapping unless it is shown that these points carry significant Gaussian weight in the vMCG expansion. Please support the mechanistic conclusions with GBF-population-weighted geometrical distributions, and similarly qualify the two-route interpretation based on Figures S13 and S14.","section":"Section 3.2; Figure 4"},{"comment":"The claim that CH stretch modes are crucial for successful isomerisation is based on the observation that the 27-mode calculation (which excludes CH stretches) does not relax to the cis structure. This comparison is not fully controlled: the reduced-dimensionality runs use independent databases, freeze several modes that are active in the full run (for example q16, q23, and q27 appear in the active-mode analysis of Section 3.1), and removing modes also removes their zero-point energy and changes the local harmonic approximation. The 13- and 19-mode calculations exclude many other modes as well, so they do not isolate CH stretches. A more direct test would be to freeze only the CH-stretch modes while keeping all other modes active, ideally using the same database, before drawing the mechanistic conclusion.","section":"Section 3.3; Figure 6"}],"minor_comments":[{"comment":"The text refers to 'equation 3' when introducing the vMCG ansatz, but the ansatz is equation (1); also, the local harmonic expansion in equation (3) omits the remainder term and does not define the notation for the Hessian.","section":"Section 2.2"},{"comment":"The x-axis of Figure 5 is labeled in arbitrary units and the 'vector that connects the S0_trans and MECI_db geometries' is not defined; please describe how the one-dimensional cut was constructed and report the barrier height in a well-defined coordinate.","section":"Figure 5"},{"comment":"Reference 39 is incomplete: it gives only 'J. Chem. Phys., 2016, 144, 024111' without authors or title, which makes it difficult for readers to identify the coupled-coherent-states paper being cited.","section":"References"},{"comment":"The text mentions '19 and 24mode' calculations, but the preceding sentence defines only 13, 19, and 27 modes; please correct this typo and verify the number of modes in each reduced-dimensionality run.","section":"Section 3.3"},{"comment":"The data-availability statement lists the 20GWP and 16GWP databases but does not specify the versions of Molpro and Quantics used or the convergence thresholds for the SA(2)-CASSCF energies, gradients, and nonadiabatic couplings; adding these details would improve reproducibility.","section":"Data availability"}],"recommendation":"major_revision","confidential_remarks":"The paper's main selling point is the comparison between DD-vMCG and established trajectory methods, but the 'same electronic structure method' claim in the abstract is not accurate for the cited benchmarks. I would encourage the authors to reframe the contribution as a demonstration of DD-vMCG's capability on a full-dimensional retinal model and to present the delayed timescale as a method- and surface-dependent observation, rather than as a direct method benchmark. If the authors can add a trajectory calculation on the identical SA(2)-CAS(6,6)/6-31G surface, or otherwise quantify the electronic-structure differences, the paper would be considerably stronger."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is the first DD-vMCG simulation of trans-PSB3 in full 36-dimensional space. That is a nontrivial computation, and the raw result is plausible: the average torsional angle reaches cis values by 250–300 fs, population transfer lags the AIMS/TSH benchmarks, and the reduced-dimensionality runs show that CH stretch modes are required for isomerization. The authors also deposit the full dynamics database on Zenodo, which is real and checkable evidence. I credit that.\n\nThe paper's central claim, though, is not as clean as the abstract makes it. The abstract says 'the same electronic structure method' is used in the comparison, but Section 2.1 is explicit that the Martinez AIMS work used a three-state average and Szymczak et al. added MRCI after SA(2)-CASSCF, while this work uses plain SA(2)-CAS(6,6)/6-31G. So the delay could be caused by the missing S2 state or missing dynamic correlation rather than by DD-vMCG itself. That distinction matters because the whole selling point is that quantum wavepacket dynamics gives a qualitatively different timescale. The confound is addressable—run a trajectory method on the same SA(2)-CAS(6,6)/6-31G surface—but as presented the comparison is method-plus-surface, not method-only. This is the softest spot, and it is moderately soft, not fatal.\n\nTwo smaller issues. Figure 5, the barrier invoked as the cause of the delay, has arbitrary-unit axes and no energy values, so it cannot independently support the mechanism. And the system still has 60% S1 population at 300 fs, meaning the 'fully isomerised' statement rests entirely on geometry, not electronic state. The authors should say that more plainly.\n\nOn the positive side, the mode-by-mode analysis is careful, the CH-stretch dependence is a concrete new finding, and the paper does not oversell the method beyond what the data show. The self-citations to previous vMCG/MCTDH studies are relevant context, not padding. The 20-GBF convergence is in the SI, though the main text would benefit from a one-line summary of it.\n\nWho is this for? Nonadiabatic dynamics method developers and people benchmarking PSB3. It deserves a serious referee: the application is legitimate, the data are there, and the confound is fixable in revision. I would send it out, but with a clear request to correct the abstract and either run the controlled trajectory comparison or restrain the claims to what the current evidence supports.","headline":"First DD-vMCG run on PSB3 shows a genuinely delayed isomerization, but the headline method-vs-trajectory comparison is confounded by different electronic structure levels; the abstract overclaims.","tokens_in":12394,"tokens_out":1452,"would_cite":true,"duration_ms":15385,"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":"The paper claims that a full-dimensional quantum wavepacket method, DD-vMCG, makes the benchmark retinal model trans-PSB3 isomerise roughly 200-250 femtoseconds slower than trajectory methods, with the delay caused by a barrier before the…","keywords":["photoisomerisation","PSB3","DD-vMCG","conical intersection","quantum wavepacket dynamics","nonadiabatic dynamics","retinal protonated Schiff base","CASSCF"],"falsifier":"Repeat the benchmark surface-hopping and ab initio multiple spawning simulations using exactly the same interpolated two-state surface stored in the paper's database; if those trajectory methods also show a delay of roughly 200 fs, the delay is a property of the surface, not of quantum wavepacket dynamics.","tokens_in":11295,"feed_emoji":"⚛","tokens_out":13020,"duration_ms":119967,"temperature":0.7,"pith_summary":"The paper applies the direct-dynamics variational multiconfigurational Gaussian (DD-vMCG) wavepacket method, in full 36-dimensional space, to the benchmark retinal model trans-PSB3, and claims that fully quantum nuclear dynamics makes the trans-cis photoisomerisation hundreds of femtoseconds slower than trajectory-based methods report. The delay is attributed to the accessibility of the conical intersection: the wavepacket first becomes trapped at the $S_1$ minimum of the trans form, then crosses an energy barrier before population transfers to the ground state. In the simulations, the isomerisation completes by 250-300 fs, significant $S_1$-to-$S_0$ transfer begins only after 200 fs, and the process fails when the CH stretch modes are frozen. If true, this means the choice of nuclear-dynamics method can change the apparent timescale and mechanism of a textbook photochemical reaction, which matters for how retinal photoisomerisation is modelled and benchmarked.","feed_headline":"200 fs delay: quantum wavepacket lags trajectories in retinal model","feed_subtitle":"The delay comes from a barrier before the conical intersection that trajectory methods cross quickly.","key_machinery":"The load-bearing object is the DD-vMCG wavefunction ansatz, a variational multiconfigurational Gaussian wavepacket whose centers and momenta are propagated by the Dirac-Frenkel principle instead of by classical trajectories. On-the-fly electronic structure supplies energies, gradients, Hessians, and nonadiabatic couplings; propagation diabatisation converts the adiabatic surfaces into smooth diabatic ones to avoid geometric phase issues, and Shepard interpolation over a growing database of 48,857 electronic-structure points makes the propagation affordable. The local harmonic approximation around each Gaussian center supplies the potential integrals needed for a variational propagation without a precomputed global grid.","core_discovery":"The central discovery is that DD-vMCG with SA(2)-CAS(6,6)/6-31G gives a qualitatively different deactivation timescale for PSB3 than independent-trajectory methods: the $S_1$ population starts transferring to $S_0$ only after about 200 fs and sits at a 60:40 mixture of $S_1$ and $S_0$ at 300 fs, even though the torsional angle has moved from about 180 degrees to 0 degrees and the BLA parameter has stabilised near 1.4 \\AA. A database of 48,857 electronic-structure points shows the wavepacket lingering at the $S_1$-trans minimum, with only 0.4% of stored structures near the lowest-energy conical intersection, and the energy profile along $S_1$ shows a barrier before that intersection. The dynamics follow a sequential route: bond-length stretching in the first 25 fs, torsional activation only after about 150 fs, population transfer after 200 fs, and CH stretch modes dominating after 220 fs and driving the system to the cis minimum. Reduced-dimensionality runs that exclude CH stretches never reach the cis geometry, which the paper reads as evidence that those modes are required for successful isomerisation.","pith_inferences":["Beyond the paper: because the earlier trajectory benchmarks used three-state averaging or added MRCI corrections, part of the reported delay could come from the simpler two-state CASSCF surface rather than from quantum wavepacket dynamics; a method-only comparison cannot separate these contributions.","Beyond the paper: the variational coupling between Gaussian centres maintains quantum coherence across all 36 modes, so the lingering $S_1$ population may reflect coherent wavepacket motion that trajectory methods average away; the paper shows time-dependent mode widths but does not quantify coherence loss.","Beyond the paper: an isotopic test follows immediately, deuterating the CH bonds should shift the timing or yield of isomerisation if those stretches are the gating modes, and the same dynamics could be rerun with deuterium masses."],"forward_implications":["PSB3 photoisomerisation timescales are method-dependent: trajectory-based values near 60 fs and wavepacket values near 250-300 fs are both possible outputs, so benchmark comparisons need to specify the nuclear-dynamics method.","A barrier before the conical intersection becomes a concrete dynamical feature: the wavepacket lingers at the $S_1$-trans minimum, so transient trapping should be observable in time-resolved measurements if this picture holds.","CH stretch modes gate the isomerisation: reduced-dimensional models that omit them will not reproduce the cis product, so effective-mode models of retinal chromophores should keep the high-frequency stretches.","Two deactivation routes are available, one through a 120-degree conical intersection and one through an approx 80-degree near-degenerate region, and the wavepacket chooses between them depending on the shape of the surface.","Full-dimensional on-the-fly quantum dynamics is now feasible for a 36-mode benchmark system, so similar wavepacket analyses can be applied to other photoactive molecules."],"supporting_citations":[{"why":"Supplies the ab initio multiple spawning benchmark with three-state averaging against which the DD-vMCG timescales are compared.","marker":"25"},{"why":"Supplies the surface-hopping benchmark showing near-complete transfer on the timescale the paper contrasts with its own.","marker":"27"},{"why":"Supplies the earlier SA(2)-CASSCF(6,6) plus MRCI reference surface that this work's plain CASSCF surface is meant to approximate.","marker":"32"},{"why":"Supports the choice of the 6-31G basis by showing it reproduces vertical excitation energies, conical intersection locations, and reaction paths compared with polarized basis sets.","marker":"34"},{"why":"Provides a previous potential-energy-surface description of PSB3 with the same active space and basis set, anchoring the model setup.","marker":"9"},{"why":"Supplies the variational Gaussian-wavepacket equations of motion that define the vMCG ansatz used here.","marker":"37"},{"why":"Introduces the on-the-fly direct-dynamics variant of vMCG that makes the 36-mode propagation possible.","marker":"41"},{"why":"Describes the propagation diabatisation used to build smooth diabatic surfaces near the conical intersection.","marker":"42"}],"fun_headline_variants":["Quantum wavepacket idles 200 fs where trajectories sprint","Retinal model: quantum dynamics slow isomerisation by 200 fs","Wavepacket stalls at barrier, CH stretches unlock cis form","DD-vMCG reveals hidden barrier in retinal photoisomerisation","Quantum delay: CH stretch modes are the key to retinal twist"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's central comparison assumes that the simpler two-state electronic-structure surface used here is close enough to the three-state-averaged and MRCI-corrected surfaces used in the benchmark trajectories that the observed delay is attributable to the quantum dynamics method, not to the missing second excited state or missing dynamic electron correlation.","fun_headline_variants_meta":{"raw":{"variants":["Quantum wavepacket idles 200 fs where trajectories sprint","Retinal model: quantum dynamics slow isomerisation by 200 fs","Wavepacket stalls at barrier, CH stretches unlock cis form","DD-vMCG reveals hidden barrier in retinal photoisomerisation","Quantum delay: CH stretch modes are the key to retinal twist"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000203,"raw_usage":{"total_tokens":1392,"prompt_tokens":957,"completion_tokens":435,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":573,"completion_tokens_details":{"reasoning_tokens":347}},"tokens_in":573,"tokens_out":435,"duration_ms":4429,"temperature":1.0,"reasoning_tokens":347,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:54:35.893616+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Repeat the benchmark surface-hopping and ab initio multiple spawning simulations using exactly the same interpolated two-state surface stored in the paper's database; if those trajectory methods also show a delay of roughly 200 fs, the delay is a property of the surface, not of quantum wavepacket dynamics.","supporting_citations":[],"review_version":1}