{"id":"3cf9fe49-cb2f-431d-8097-a1680b4ade88","arxiv_id":"2611.16780","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"In a two-site model with quadratic electron-phonon coupling, four gradient-optimized drive parameterizations (piecewise-constant, B-spline, feed-forward, recurrent) raise time-averaged double occupation by 0.074 to 0.200 over a monochromatic reference.","lead":"Gradient-based optimization through an exact simulation of a two-site electron-phonon model finds drive shapes that raise the time-averaged double occupation by 11 to 30 percent over a simple periodic drive. It is an explicitly labeled proof of concept, but it is a concrete template for machine-designed light pulses aimed at correlated materials.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"The headline comparison lacks a fluence- or amplitude-matched baseline, so the claimed 30% improvement may reflect extra drive power rather than pulse shaping.","rationale":"The reader's weakest_assumption correctly identifies the missing fluence/amplitude constraint as the key weakness. The paper's own concession (b118) confirms this. My concern is identical: the comparison is not apples-to-apples, so the headline 30% improvement may be an artifact of unequal drive power. The paper is otherwise sound and transparent about limitations, so the verdict should remain CONDITIONAL—the condition being that the fluence-matched comparison is performed. I do not recommend REJECT because the paper's qualitative claim (shaped fields can raise D) is supported, and the quantitative claim is explicitly conditional. The concrete test I propose would settle the issue directly. I agree with the reader's assessment.","tokens_in":10011,"tokens_out":562,"duration_ms":8674,"concrete_test":"Re-run the optimization of the B-spline, vector, and RNN controllers with a penalty on the time-integrated squared drive (fluence) equal to that of the monochromatic reference, or with a hard cap on peak amplitude. If the B-spline's mean D drops to within 0.02 of the reference, the claimed advantage of structured fields is not supported. If it remains above 0.80, the shaping effect is real. Report the reference amplitude and frequency explicitly.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central quantitative claim—that the B-spline controller raises mean double occupation by ~30% relative to the monochromatic reference—is undermined because the reference drive's frequency and amplitude are never stated, and the B-spline field visibly has larger excursions (Fig. 3b). The author explicitly concedes this in Sec. V.A: 'A further comparison at fixed peak amplitude, integrated squared drive, bandwidth, and optimization budget would isolate the contribution of the representation itself' (b118). Without such a constraint, the comparison conflates the effect of pulse shaping with the effect of simply using a stronger or differently tuned drive. The ranking in Table I is therefore uninterpretable as evidence that structured fields are inherently superior; it only shows that the optimizer found a field that works better than the chosen monochromatic one, which may be suboptimal in amplitude or frequency. This is a load-bearing concern because the paper's headline result (abstract, b4) and the strongest claim (reader's summary) both rest on this comparison. The paper acknowledges the limitation but does not resolve it, leaving the central claim conditional on a control that has not been performed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"The paper presents a differentiable control framework for a two-site electron-phonon model with quadratic coupling, comparing four parameterizations (vector, B-spline, feed-forward, recurrent) to maximize time-averaged double occupation. The central claim is that pulse shaping can substantially increase double occupation relative to a monochromatic reference, with the B-spline achieving ~30% improvement. A separate recurrent network is trained to predict next-step observables. The work is presented as a computational proof of concept.","tokens_in":10256,"tokens_out":1565,"duration_ms":15075,"significance":"If the central claim holds, the work demonstrates that structured, non-periodic drives can outperform simple monochromatic drives in enhancing electronic correlations via nonlinear electron-phonon coupling, which could inform experimental efforts in light-driven materials control. The use of differentiable optimization and neural-network predictors in this context is methodologically interesting. However, the significance is tempered by the minimal model (two sites, zero Hubbard U, no dissipation) and the acknowledged lack of fair comparison baselines.","major_comments":[{"comment":"The headline quantitative claim—that the B-spline controller raises mean double occupation by ~30% relative to the monochromatic reference—is not supported by a controlled comparison. The monochromatic reference's frequency and amplitude are never stated (Sec. II through V, Fig. 3 caption). The B-spline field in Fig. 3(b) visibly has larger excursions, and no constraint on peak amplitude, integrated squared drive, or bandwidth is imposed on any controller. The authors explicitly acknowledge this limitation in §V.A: 'A further comparison at fixed peak amplitude, integrated squared drive, bandwidth, and optimization budget would isolate the contribution of the representation itself' (b118). Because the comparison conflates pulse shaping with drive strength/tuning, the ranking in Table I is uninterpretable as evidence that structured fields are inherently superior. This is a load-bearing issue for the paper's central claim.","section":"§V.A, Table I, Fig. 3","authors_acknowledge":"A further comparison at fixed peak amplitude, integrated squared drive, bandwidth, and optimization budget would isolate the contribution of the representation itself.","authors_acknowledge_p":0.94},{"comment":"The results in Table I are reported as single trajectory values without uncertainty estimates across training runs or random initializations. The authors note this in b131: 'These are individual trajectory results, without uncertainty estimates across training runs.' For a study that compares four control parameterizations and draws conclusions about their relative performance, the absence of any statistical characterization (e.g., mean and standard deviation over multiple optimizations with different seeds) makes it impossible to assess whether the observed differences are significant or reproducible. This is particularly concerning given the acknowledged sensitivity to initialization and optimization details.","section":"§V.A, Table I","authors_acknowledge":"These are individual trajectory results, without uncertainty estimates across training runs.","authors_acknowledge_p":0.97},{"comment":"The predictor evaluation uses teacher forcing (simulator observables fed at every step) rather than free-running rollout. The authors correctly state this in §V.B: 'The comparison therefore evaluates teacher-forced one-step prediction, rather than a free-running 1000-step rollout.' However, the paper's framing in the abstract and conclusion suggests a learned dynamics model suitable for control, which would require stable multi-step prediction. The reported errors (e.g., MAE/range 4.9% for double occupation) are only for one-step conditioned predictions and do not demonstrate that the learned model can be reliably used for closed-loop control or simulation beyond a single step. The paper should either present free-running results or clearly delimit the scope of the predictor's demonstrated capability.","section":"§V.B, Table II, Fig. 5","authors_acknowledge":null,"authors_acknowledge_p":null}],"minor_comments":[{"comment":"The abstract states 'all the methods increase double occupation, with the B-spline raising the mean total double occupation of approximately 30%.' This is ambiguous: it could be read as 30 percentage points or 30% relative increase. From Table I, 0.200/0.655 ≈ 30.5%, so it is relative. Clarify in the abstract.","section":"Abstract, §I","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The Hamiltonian uses ˆneli = Pσ ˆc†iσˆciσ, but the notation ˆneli↑ ˆneli↓ in the Hubbard term is inconsistent with the definition of ˆneli. It should be ˆneli↑ ˆneli↓ or (ˆneli↑)(ˆneli↓). Minor.","section":"§II, Eq. (1)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The B-spline basis functions are shown over 0 ≤ t ≤ 50, but the control duration in Fig. 3 is tΩ = 25 (i.e., t = 50 for Ω=0.5). Clarify the relationship between time units and Ω in the figure caption.","section":"§III.C, Fig. 2","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The loss function weights double occupation by a factor of 10. The rationale for this specific choice is not discussed. Since the predictor is a separate component, the sensitivity of the learned dynamics to this weight should at least be mentioned.","section":"§IV.C, Eq. (13)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The FNN training progression in Fig. 4 shows improvement from initialization to 50 epochs. However, the FNN trajectory in Fig. 3 is only shown for tΩ = 0–10, and its mean (0.784) is compared to the monochromatic reference over a different interval. The caption notes this, but the main text could be clearer that the FNN and other methods are not directly comparable due to different intervals.","section":"§V.A, Fig. 4","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The MAE/range normalization is described as 'a descriptive normalization, not a confidence interval or a pointwise relative error.' This is appropriate, but the range values (0.456, 8.680, 5.470) should be explicitly stated in the table or caption for reproducibility.","section":"§V.B, Table II","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The conclusion mentions 'suggesting directions to improve control, induced and sustained superconductivity or other non-equilibrium phases of matter.' This overstates the implications given the minimal model and the lack of controlled comparison. Tone down or qualify.","section":"§VI","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The paper uses 'double occupation' to mean the probability that both electrons occupy the same site (summed over sites). This is fine, but note that in the Hubbard model literature, 'double occupancy' often refers to the site-averaged quantity. Clarify early to avoid confusion.","section":"General","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"The paper does what it says: it sets up a differentiable control loop for a two-site electron-phonon model, optimizes four drive representations against accumulated double occupation, and trains a separate recurrent predictor. The machinery is sound, the code presumably runs, and the author flags most of the obvious caveats—no error bars, teacher-forced evaluation, missing fluence constraint. For a proof of concept on a 1600-dimensional Hilbert space, that is fine.\n\nThe soft spot is the one the stress-test note hits. The headline number—B-spline raises mean double occupation by ~30% relative to a monochromatic reference—is the paper's strongest claim and it is uninterpretable as stated. The reference drive's frequency and amplitude are never given. The B-spline field visibly has larger excursions. The author concedes in Sec. V.A that a fixed-amplitude, fixed-fluence comparison would be needed to isolate the representation effect. Until that control is done, Table I shows that the optimizer found a field that works better than the chosen monochromatic one, but it does not show that pulse shaping per se is responsible for the gain. The ranking across representations is also uninterpretable under this confound.\n\nThe predictor result is modest: teacher-forced one-step MAE of 4.9% range-normalized for double occupation, on one sequence, with no free-running rollout. The paper identifies the rollout as the missing step toward the Fig. 1(b) architecture, so this is an acknowledged limitation, not a hidden flaw.\n\nWho is this for? Someone who wants a quick template for differentiable control in a small quantum model and is willing to redo the comparison properly. The paper is not wrong; it is incomplete in a way that the author acknowledges. A serious referee would ask for the fluence-matched comparison and a free-running predictor evaluation before publication. I would not cite it in its current form, but I would not desk-reject it either—the core idea is worth a revision.\n\nRecommendation: conditional acceptance after the author supplies the missing baseline comparison and, ideally, a free-running rollout of the predictor. The paper is honest and the code is presumably available; the fix is well-defined.","headline":"Honest two-site demo with a load-bearing comparison gap; the 30% improvement claim is uninterpretable without a fluence-matched baseline.","tokens_in":10947,"tokens_out":539,"duration_ms":7221},"pith_extraction":{"msc":[],"pacs":["71.38.-k","78.20.Bh","02.30.Yy","07.05.Mh"],"model":"deepseek/deepseek-v4-flash","headline":"Differentiable control shapes light-driven electron-phonon dynamics to boost pairing","keywords":["electron-phonon coupling","quadratic coupling","optimal control","automatic differentiation","B-spline","recurrent neural network","double occupation","nonequilibrium dynamics"],"falsifier":"If the B-spline optimization is repeated with a constraint that the integrated squared drive (fluence) is held equal to the monochromatic reference, and the resulting double occupation increase drops below 10% or vanishes, then the central claim that pulse shaping substantially enhances pairing would be falsified, and the advantage would be attributable to increased drive strength.","tokens_in":11313,"feed_emoji":"🔬","tokens_out":1667,"duration_ms":14611,"temperature":0.7,"pith_summary":"This paper demonstrates that gradient-based optimization of a phonon drive can substantially increase electronic double occupation in a minimal two-site model with quadratic electron-phonon coupling, compared to a monochromatic laser reference. The author compares four control representations—piecewise-constant vector, B-spline expansion, feed-forward neural network, and recurrent neural network—all optimized through automatic differentiation of the simulated quantum dynamics. The B-spline representation achieves the largest increase, raising the time-averaged double occupation by about 30% relative to the reference. A separate recurrent network learns to predict the next-step observables from the current state and drive, with small one-step errors. The results establish a computational proof of concept for shaping local electronic correlations via structured phonon drives, with implications for nonequilibrium control of materials.","feed_headline":"Optimized phonon drive boosts electronic pairing by 30%","feed_subtitle":"Differentiable control of a two-site electron-phonon model shows that shaped fields outperform monochromatic lasers","key_machinery":"The central machinery is the differentiable optimization loop that propagates gradients through the simulated quantum dynamics of the Hamiltonian H(t) = H_0 + α(t) Σ_i X_i, with quadratic electron-phonon coupling g_2 Σ_i n_i $X_i^{2}$. Automatic differentiation through the discretized time evolution (with time step δt = 0.1) allows gradient-based maximization of the accumulated double occupation reward R(θ) = Σ_n D_n. Four control parameterizations are compared: a vector of independent amplitudes (M parameters), a B-spline expansion with 20 degree-10 basis functions, a feed-forward neural network mapping observables to field, and a recurrent GRU network with hidden state. The quadratic coupling term is the physical mechanism that makes the phonon potential occupation-dependent, enabling control of electronic correlations via the lattice drive.","core_discovery":"The central claim is that differentiable optimization of a phonon drive, parameterized in a compact B-spline basis, can raise the time-averaged electronic double occupation by approximately 30% relative to a monochromatic reference in a two-site quadratic-coupling electron-phonon model at zero Hubbard interaction. The B-spline trajectory maintains elevated double occupation over much of the interval, showing that structured, non-periodic fields exploit the coupled dynamics more effectively than a simple sinusoidal drive. The recurrent controller also achieves gains and can be evaluated beyond its training horizon, while a separate recurrent network predicts the next-step observables with mean absolute errors of 0.022 for double occupation, 0.163 for phonon number, and 0.038 for phonon coordinate (range-normalized 4.9%, 1.9%, 0.69%). The author frames this as a computational proof of concept for control and prediction in a minimal model, suggesting a route toward learned models for nonequilibrium materials control.","pith_inferences":["The paper leaves open the crucial question of whether the B-spline gain is primarily due to higher integrated drive power rather than temporal structure; the author explicitly acknowledges that a comparison at fixed peak amplitude, integrated squared drive, and bandwidth is needed. A reader should interpret the 30% improvement as conditional on the chosen monochromatic reference, not as a universa","The two-site model at U=0, while tractable, omits the Hubbard interaction that is central to many correlated materials. Extending the optimization to finite U and including dissipation would test whether the control strategies remain effective when electronic correlations compete with the phonon-mediated pairing.","The recurrent predictor's teacher-forced evaluation (using simulator observables as inputs) does not test stability under free-running rollout; coupling the predictor to the controller in a closed loop (as sketched in Fig. 1b) would reveal whether prediction errors accumulate and degrade control performance."],"forward_implications":["If the B-spline gain is robust under fixed drive constraints, then compact basis representations may generally outperform direct parameterization for quantum control of coupled electron-phonon systems.","The recurrent controller's ability to operate beyond its training horizon suggests that learned feedback strategies could be deployed in experimental settings where the drive must adapt to unmodeled drift or noise.","The separate dynamics predictor, even with small one-step errors, could be used in a model-predictive control loop that replaces the full quantum simulator, enabling optimization over longer timescales or in larger systems.","The approach may generalize to other observables (e.g., current, magnetization, or superconducting order parameters) by changing the reward function, providing a flexible platform for nonequilibrium materials control.","Extending the two-site proof of concept to periodic lattices with many phonon modes could test whether the shaped-drive advantage survives in systems with competing orders and dissipation."],"supporting_citations":[{"why":"Sentef 2017: provides the two-site electron-phonon model with quadratic coupling that the paper adopts as its physical system.","marker":"[14]"},{"why":"Puviani and Sentef 2018: justifies the quadratic coupling as the symmetry-allowed nonlinear phonon-electron interaction, which is the physical mechanism for control.","marker":"[15]"},{"why":"Khaneja et al. 2005 (GRAPE): supplies the gradient-based pulse engineering methodology that the differentiable control approach builds on.","marker":"[16]"},{"why":"Leung et al. 2017: demonstrates automatic differentiation through quantum simulation for control, which the paper uses as its core optimization technique.","marker":"[17]"},{"why":"Porotti et al. 2023: provides a recent example of differentiable quantum control with neural networks, supporting the use of automatic differentiation.","marker":"[19]"},{"why":"Kingma and Ba 2015 (Adam): the optimization algorithm used for updating control parameters.","marker":"[25]"},{"why":"de Boor 2001: the standard reference for B-spline basis functions, which the paper uses for compact drive parameterization.","marker":"[26]"},{"why":"Cho et al. 2014 (GRU): defines the gated recurrent unit architecture used for the recurrent controller and predictor.","marker":"[28]"}],"fun_headline_variants":["Shaped phonon drives lift electronic pairing 30%","Optimized phonon fields boost electron double occupation","Differentiable control tunes electron-phonon dynamics","Neural-learned phonon drives gain 30% in pairing","B-spline phonon drive outperforms monochromatic pulses"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The comparison assumes that the monochromatic reference drive and the optimized drives are fairly compared, but the reference's frequency and amplitude are never stated, and the B-spline field has visibly larger excursions, so the claimed 30% improvement may partly reflect higher drive power rather than pulse shaping itself.","fun_headline_variants_meta":{"raw":{"variants":["Shaped phonon drives lift electronic pairing 30%","Optimized phonon fields boost electron double occupation","Differentiable control tunes electron-phonon dynamics","Neural-learned phonon drives gain 30% in pairing","B-spline phonon drive outperforms monochromatic pulses"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":0.00017,"raw_usage":{"cost":0.00016996,"is_byok":false,"cost_details":{"upstream_inference_cost":0.00016996,"upstream_inference_prompt_cost":0.00012712,"upstream_inference_completions_cost":4.284e-05},"total_tokens":1061,"prompt_tokens":908,"completion_tokens":153,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":0,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":72}},"tokens_in":908,"tokens_out":153,"cost_source":"openrouter_usage","duration_ms":2768,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":72,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-11-13T13:03:13.083629+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"If the B-spline optimization is repeated with a constraint that the integrated squared drive (fluence) is held equal to the monochromatic reference, and the resulting double occupation increase drops below 10% or vanishes, then the central claim that pulse shaping substantially enhances pairing would be falsified, and the advantage would be attributable to increased drive strength.","supporting_citations":[{"cited_title":"Phys. Rev. B95, 205111","cited_arxiv_id":null,"evidence_quote":"Sentef 2017: provides the two-site electron-phonon model with quadratic coupling that the paper adopts as its physical system."},{"cited_title":"Phys. Rev. B98, 165138","cited_arxiv_id":null,"evidence_quote":"Puviani and Sentef 2018: justifies the quadratic coupling as the symmetry-allowed nonlinear phonon-electron interaction, which is the physical mechanism for control."},{"cited_title":"Journal of Magnetic Resonance172, 296","cited_arxiv_id":null,"evidence_quote":"Khaneja et al. 2005 (GRAPE): supplies the gradient-based pulse engineering methodology that the differentiable control approach builds on."},{"cited_title":"Phys. Rev. A95, 042318","cited_arxiv_id":null,"evidence_quote":"Leung et al. 2017: demonstrates automatic differentiation through quantum simulation for control, which the paper uses as its core optimization technique."},{"cited_title":"PRX Quantum 4, 030305","cited_arxiv_id":null,"evidence_quote":"Porotti et al. 2023: provides a recent example of differentiable quantum control with neural networks, supporting the use of automatic differentiation."},{"cited_title":"Adam: A Method for Stochastic Optimization","cited_arxiv_id":"1412.6980","evidence_quote":"Kingma and Ba 2015 (Adam): the optimization algorithm used for updating control parameters."},{"cited_title":"de Boor,A Practical Guide to Splines, revised ed., Applied Mathematical Sciences, Vol","cited_arxiv_id":null,"evidence_quote":"de Boor 2001: the standard reference for B-spline basis functions, which the paper uses for compact drive parameterization."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Cho et al. 2014 (GRU): defines the gated recurrent unit architecture used for the recurrent controller and predictor."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}