{"id":"bc8aeaf0-9106-4e8a-86cc-0bc231d79509","arxiv_id":"2507.16349","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A U-Net trained on solver trajectories accelerates the energy-adaptive Riemannian conjugate gradient method for rotating Gross-Pitaevskii ground states, saving about 22% of iterations and 14.5% of wall time on average.","lead":"This paper combines a U-Net neural network with an existing iterative solver to speed up ground-state calculations for rotating Bose-Einstein condensates. In numerical tests the hybrid method cut iteration counts by about 22% and wall-clock time by about 14.5% on average, though it sometimes fails or hurts performance.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline speed-up averages pool same-minimum and different-minimum cases, which the authors themselves say can skew results, so the 22.06% iteration saving may not reflect like-for-like acceleration.","rationale":"The paper is an honest empirical contribution, and the strongest claim is the measured speed-up. I looked first at the acceptance heuristic flagged by the reader (Section 4.4). While it is indeed unproven, the random-application baseline in Section 5.2 provides an empirical control: the proposed acceptance/timing strategy beats random application by a wide margin (22.06% vs 5.52% mean iteration saving), so the heuristic has indirect support for the specific benchmark distribution. The acceptance criterion is a robustness and generality concern, not an immediate invalidation of the reported numbers. The more load-bearing issue is that the headline numbers pool cases the authors themselves say can skew results. Figure 8(b) reports means over all 500 test cases, coloring but not separating same-minimum vs different-minimum convergences, and the discussion indicates 12 classical non-convergences are included in some counts. In different-minima cases, saving iterations is not a like-for-like measure of acceleration: the enhanced run solves a different problem by converging to a different local minimum, and the density-error metric against the classical reference is not a meaningful error there. Section 5.4's decision to exclude these cases for the detailed iteration analysis because they can 'skew the result' makes this a clear statistical validity risk. The concrete test of recomputing blue-only means would settle whether the central claim survives. Because this is a fixable reporting and analysis issue rather than a fundamental flaw, the conditional verdict remains appropriate.","tokens_in":12706,"tokens_out":6661,"duration_ms":77319,"concrete_test":"Re-run the benchmark analysis on the 488 cases where the classical method converged, split by whether the enhanced run converged to the same local minimum/energy as the classical run (blue vs orange in Figure 8). Report mean and median iteration saving and wall-time saving for the blue-only subset, and also recompute excluding the 12 non-converged classical cases. If the blue-only mean is materially below 22.06% (say below 15%), the headline speed-up is inflated by different-minima cases; if it remains near 22%, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim (Section 5.2: mean 22.06% iteration saving, 14.51% wall-time saving, 46.20% density-error improvement over 'all 500 test cases') is computed on a mixed population: Figures 8 and 9 color blue cases where the accelerated and classical EARCG converge to the same local minimum and orange cases of different minima, yet the reported means pool both. Section 5.4 explicitly excludes different-minima examples from the best/worst iteration analysis because 'local convergence behavior may vary significantly, skewing the result.' The same section notes that this exclusion removes some subpar accelerated cases, but the headline number is not similarly restricted. Additionally, Figures 8 and 9 are described as 'evaluated on all 500 test cases', which includes 12 runs where the classical method did not converge within 30000 iterations (Section 5.3); for those runs the baseline iteration count is a stopping cap, not a converged cost. In different-minima or non-converged-classical cases the comparison is not like-for-like: the accelerated run may reach a different local minimum faster (e.g., Figure 15, with lower energy but 67.7% 'density deterioration'), so counting this as a speed-up conflates changing the problem with accelerating it. No same-minima-only mean is reported, leaving the headline number potentially inflated by the very cases the authors say skew results.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a neural-network accelerator for the energy-adaptive Riemannian conjugate gradient (EARCG) method applied to ground-state computations of the rotating Gross-Pitaevskii equation. A U-Net is trained offline on EARCG trajectories to map mid-iteration states and gradients to the solver's converged state, and it is inserted once during a prescribed acceleration window. The acceptance of the network output is decided by the L2 normalization error of the prediction, and the algorithm then continues with EARCG to the final tolerance. Numerical experiments on 500 test cases report mean reductions of 22.06% in iteration count, 14.51% in wall-clock time, and 46.20% in density error, together with histograms, best/worst case studies, and a comparison against random application of the network.","tokens_in":13000,"tokens_out":6916,"duration_ms":63846,"significance":"If the reported gains are robust, this is a useful hybrid contribution to nonlinear eigenvalue solvers, complementing end-to-end neural approaches by retaining the classical solver's convergence guarantees while reducing cost in challenging regimes. Strengths include the reproducible open-source repository, a large test set of 500 cases, explicit disclosure of failure cases, and a sensible baseline comparison against random network application. The main weaknesses are statistical: the headline means pool runs that are not like-for-like (different local minima and non-converged classical baselines), and the acceptance heuristic is asserted but not independently validated. These issues are fixable in revision, and the paper's core idea deserves further consideration.","major_comments":[{"comment":"The headline improvements of 22.06% mean iteration saving and 14.51% mean wall-time saving are computed over all 500 test cases, as stated in Section 5.2 (Figure 8b and Figure 9b), yet the histograms color-code cases where the accelerated and classical runs converge to different local minima (orange). Section 5.4 explicitly excludes such cases from the iteration best/worst analysis because 'local convergence behavior may vary significantly, skewing the result,' which applies equally to the aggregate means. The pooled number also includes 12 runs where the classical method did not converge within 30000 iterations; for those runs the baseline is a stopping cap rather than a converged cost (Figure 19). For a like-for-like acceleration claim, report the mean and median savings restricted to runs where both methods converge to the same local minimum, and report the non-converged-classical cases separately. As written, the 22.06% figure conflates acceleration with changing the target solution (for example, Figure 15 reports a 61.57% iteration saving while the density deteriorates by 67.7%).","section":"Section 5.2 and 5.4"},{"comment":"In the best-iteration examples (Section 5.4), the text states that 'in all cases, the classical and enhanced algorithm eventually converged to the same local minimum,' but Figure 19 is included as a best case even though its caption states the classical algorithm 'did not converge within 30000 steps' and the saving is quoted as 98.38% relative to that non-converged cap. This is internally inconsistent. A stopping cap is not an iteration count, and using it as the denominator inflates the saving. Either remove Figure 19 from the best-case analysis or provide a converged classical baseline; the claim of same-minimum convergence must then be adjusted accordingly.","section":"Section 5.4 and Figure 19"},{"comment":"The acceptance strategy 'crucially relies' on the normalization error e = |1 - ||φ̃||_{L2}| as a quality indicator for the network output (Section 4.4). No theoretical justification, calibration, or independent validation of this correlation is provided; the only support is the aggregate improvement of the full strategy over random application. Because the acceptance threshold e0 controls exactly when the accelerator intervenes, a bad proxy could reject the best proposals or accept bad ones, directly affecting the reported speed-up. Provide a validation plot (e.g., e versus the density/energy error on a hold-out set) or an ablation showing the results are stable over a range of e0, and state the proportion of proposals accepted by the threshold in the 500 test runs.","section":"Section 4.4"},{"comment":"The density-improvement measure imprρ in Section 5.2 is defined relative to ρ∗, 'the reference solution (i.e., the EARCG converged state),' which is also the target the network was trained to predict (Section 4.2). The metric therefore rewards agreement with the classical solver's local minimum rather than with the physical ground state. This is visible in Figure 15, where the accelerated run attains a lower energy than the classical result yet is reported as 67.7% density deterioration. The accuracy claim should be split into (i) agreement with the classical converged state and (ii) comparison of the final energies of the two runs; at present the density-improvement statistic is partially circular and can mislead.","section":"Section 5.2"}],"minor_comments":[{"comment":"The phrase 'promising an order-of-magnitude speedup across a broad class of applications' in the introduction is not supported by the reported average savings (22% iterations, 14.5% wall time); only a few best cases approach an order of magnitude. Qualify this claim.","section":"Abstract and Section 1"},{"comment":"Report standard deviations or confidence intervals for the mean savings; the histograms show substantial spread (e.g., for random application the mean 5.52% differs markedly from the median 12.62%).","section":"Section 5.2"},{"comment":"Define what 'correct energy' means in the sentence 'random acceleration converged to the correct energy in 439 cases'; clarify whether this refers to the same local minimum as the classical method or to a global-energy criterion.","section":"Section 5.2"},{"comment":"Beyond the major inconsistency, the caption of Figure 19 states 'no estimation of the density improvement can be given,' which contradicts the general exclusion rule in Section 5.3; cross-check all figure captions for consistency with the methodology.","section":"Section 5.4 and Figure 19"},{"comment":"The observation that the non-phase-invariant loss L(φ̂, φ) outperforms the invariant loss L̂(φ̂, φ) = ||ρ̂ - ρ||_{L1} is reported without quantitative evidence; include the training or validation losses for both losses.","section":"Section 4.3"},{"comment":"The phrase '20 log-equidistantly spaced tolerances' should be '20 logarithmically equidistant tolerances' for clarity.","section":"Section 4.2"},{"comment":"The keyword 'Riemmanian' is misspelled; it should be 'Riemannian'.","section":"Keywords"},{"comment":"Reference [30] is listed as 2025 without a journal or preprint identifier; since Algorithm 1 and the step-size strategy depend on it, provide a stable reference or summarize the relevant details in the text.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a solid empirical study with reproducible code, and the authors are candid about failure cases and limitations. The main risk is that the headline statistics overstate the speed-up by mixing non-comparable runs; this is fixable with a stratified analysis. The acceptance heuristic deserves direct validation; without it, the method's mechanism is under-supported. I would not reject, but the revision needs to address the pooling issue and the internal inconsistency around Figure 19, as well as the circularity of the density metric."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Peterseim et al. graft a U-Net onto the energy-adaptive Riemannian CG solver for rotating Gross-Pitaevskii and show it can shave iterations and wall time. The genuinely new bit is applying learned acceleration to a Riemannian eigenvalue solver, rather than to convex optimization, and using the normalization error of the un-normalized network output as a cheap acceptance criterion. That is a clever trick, and it works in their experiments: over 500 random parameter/initial-data cases they report ~22% fewer iterations and ~14.5% less wall-clock time on average, with only a few cases where the network makes things worse. They also ship the code and report failure cases, which is more than most ML-for-PDE papers do.\n\nThe main soft spot is the headline number. The 22% average pools cases where the accelerated and classical runs converge to different local minima, plus a handful of runs where the classical baseline hits the 30,000-iteration cap. The authors explicitly exclude different-minima cases from the best/worst iteration analysis in Section 5.4 because they \"skew the result,\" but they do not exclude them from the headline. Some of those cases are counted as big savings simply because the baseline is a stopping cap, or because the solver lands in a different (sometimes lower-energy) basin. The true same-minimum speed-up is likely smaller, though the median saving of ~20% suggests there is still a real effect. They should report the same-minima-only mean and median.\n\nThe acceptance heuristic -- accepting the network output only when its L2 norm is close to 1 -- is empirically motivated but has no theoretical backup. That is acceptable for a numerical methods paper if they gave evidence that it correlates with density improvement, but they don't; they just show the histograms. Relatedly, the \"density improvement\" metric compares to the classical solver's converged state, which the network was trained on, so that metric is partly circular. The independent evidence is the iteration and wall-time counts, which are not circular, so the central claim does not rest on the circular part.\n\nOverall, this is a solid empirical contribution for people working on GPE solvers or learned acceleration. It is not a breakthrough, but it is a useful step for a notoriously hard regime of rotating GPE. I would send it to peer review; the authors should be asked to add same-minima-only statistics, a comparison against a second baseline solver, and at least one out-of-distribution test to show the heuristic generalizes.","headline":"A useful empirical hybrid acceleration for a hard 2D GPE solver, with an honestly reported but partly inflated headline speed-up that needs a same-minima-only breakdown before the exact numbers should be trusted.","tokens_in":13524,"tokens_out":2438,"would_cite":true,"duration_ms":26053,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65H17","65N25","81Q10","35Q55","58C40"],"pacs":[],"model":"deepseek-v4-flash","headline":"A neural network inserted mid-iteration into the energy-adaptive Riemannian conjugate gradient solver for the rotating Gross-Pitaevskii equation cuts iteration counts by 22% on average and wall-clock time by 14.5% over 500 test cases…","keywords":["nonlinear Schrödinger eigenvalue problem","Gross-Pitaevskii equation","Bose-Einstein condensate","Riemannian optimization","energy-adaptive Riemannian conjugate gradient","neural network acceleration","U-Net","normalization error"],"falsifier":"Compute the rank correlation between the normalization error $e$ and the true density error $\\|\\tilde{\\rho} - \\rho^*\\|_{L^1}$ across the 40,000 training pairs; if the correlation is near zero, the acceptance gate is not selecting good proposals and the reported speedups would need a different explanation.","tokens_in":12483,"feed_emoji":"🌀","tokens_out":8972,"duration_ms":83975,"temperature":0.7,"pith_summary":"The paper tries to establish that a trained neural network can accelerate a classical iterative solver for nonlinear Schrödinger eigenvalue problems rather than replace it. The testbed is the rotating Gross-Pitaevskii equation for Bose-Einstein condensates, where the energy-adaptive Riemannian conjugate gradient (EARCG) method can stagnate for thousands of iterations while a vortex lattice slowly rotates into alignment with a slightly asymmetric trap. The accelerator is a U-Net that sees the current state and gradient and proposes an improved state; the proposal is accepted only when its $L^2$ normalization error is small, then renormalized and handed back to EARCG. Across 500 randomized test cases this strategy saves 22.06% of iterations and 14.51% of wall-clock time on average, and reduces density error by 46.20% on average, far better than applying the same network at a random iteration. If correct, the template extends to other Riemannian solvers used in electronic structure and multicomponent condensates.","feed_headline":"Neural net cuts BEC solver iterations by 22 percent","feed_subtitle":"A U-Net proposal inside Riemannian conjugate gradient cuts wall time 14.5% and density error 46%.","key_machinery":"The load-bearing object is the EARCG iteration on the $L^2$ unit sphere $\\mathbb{S} = \\{\\varphi \\in H^1 : \\|\\varphi\\|_{L^2} = 1\\}$, equipped with the energy-adaptive metric $g^{\\mathrm{a}}_{\\varphi}(v,w) = a_{\\varphi}(v,w)$ and the normalization retraction. The accelerator is a U-Net, a convolutional encoder-decoder originally built for image segmentation, with 31 million parameters: it takes an $(n,n,4)$ tensor holding the real and imaginary parts of the current state and the current gradient, and outputs an $(n,n,2)$ enhanced state. The mechanism that carries the argument is the acceptance gate: because the network has no normalization layer, the scalar $e = |1 - \\|\\tilde{\\varphi}\\|_{L^2}|$ serves as a cheap proxy for proposal quality, and only proposals with $e < e_0 = 5 \\times 10^{-3}$ are normalized and fed back into EARCG. This keeps the network a proposal generator inside the original solver rather than an end-to-end surrogate.","core_discovery":"The central claim is that a single, well-timed neural network proposal can bypass the slowest phase of EARCG for rotating condensates. The network is trained offline on 40,000 pairs gathered from EARCG runs: each pair contains an intermediate state and energy-adaptive gradient at a tolerance in $[10^{-4}, 10^{-1}]$ together with the converged state. When the running solver reaches that tolerance window, the U-Net proposes a new state, and the proposal is accepted only if $|1 - \\|\\tilde{\\varphi}\\|_{L^2}| < 5\\times 10^{-3}$; the accepted state is normalized and the solver continues to tolerance $10^{-8}$. On 500 test problems this reduces iterations by 22.06% on average (median 19.89%) and wall time by 14.51% (median 13.92%), with density error reduced by 46.20% on average; iteration counts improve in 91.8% of cases and wall time in 75%. The strongest gains occur in near-symmetric traps ($v_1 \\approx 1.1$, $v_2 = 1$), where the network learns to rotate the vortex lattice into the preferred orientation, cutting runs such as 25,285 iterations to 267.","pith_inferences":["One could isolate the contribution of the acceptance gate by replacing the U-Net proposals with random states whose normalization errors match the network's; if the speedup persists, the gate rather than the learned content is doing the work.","The 46% average density-error improvement suggests the network has learned the rotation-alignment step, so an architecture with built-in rotational equivariance might generalize better outside the trained parameter box.","Because the training data comes entirely from EARCG trajectories, the acceleration is tied to that solver's convergence bias; a network trained on trajectories from a different Riemannian solver may not transfer, which is a natural open question.","The reported wall-time saving is a CPU-only number; the difference between iteration savings and time savings quantifies inference overhead, and GPU inference is an obvious untested lever."],"forward_implications":["A single well-timed network proposal cuts iterations in about 92% of test cases and wall-clock time in 75%, with mean savings of 22.06% and 14.51% respectively across 500 runs.","The largest gains occur in the near-degenerate trap regime ($v_1 \\approx 1.1$, $v_2 = 1$) where EARCG stagnates; reported examples shrink runs from 25,285 to 267 iterations, a 98.94% reduction.","The normalization-error gate makes acceleration reliable: only 5.3% of runs worsen the density under the strategy, versus 40% under random application of the same network.","Because Hartree-Fock, Kohn-Sham, and multicomponent BEC problems share the same Riemannian solver structure, the same acceleration workflow transfers with minimal adaptation.","Allowing multiple accelerations per run and moving inference to a GPU are stated as likely further improvements to wall-time savings."],"supporting_citations":[{"why":"Supplies the energy-adaptive Riemannian conjugate gradient method and its implementation that the neural network accelerates.","marker":"[30]"},{"why":"The U-Net convolutional architecture on which the accelerator is based.","marker":"[31]"},{"why":"The plane-wave density-functional toolkit used to discretize and run the Gross-Pitaevskii experiments.","marker":"[24]"},{"why":"Establishes the energy-adaptive Riemannian optimization framework on the Stiefel manifold that EARCG generalizes.","marker":"[4]"},{"why":"Provides the contemporary overview of Gross-Pitaevskii numerical methods that motivates the benchmark and the comparison class.","marker":"[22]"}],"fun_headline_variants":["Neural proposal cuts BEC iterations by 22%","U-Net speeds rotating condensate solver 14.5%","NN proposal cuts BEC density error by 46%","Neural net accelerates BEC solver, cuts iterations 22%","AI proposal trims rotating BEC iterations by 22%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire acceptance strategy depends on the assumption that a small normalization error $|1 - \\|\\tilde{\\varphi}\\|_{L^2}|$ reliably means the neural network's proposal is a good approximation to the ground state; the paper gives no proof or independent validation of that correlation, so a well-normalized but physically wrong proposal would be accepted and a poorly normalized good proposal rejected.","fun_headline_variants_meta":{"raw":{"variants":["Neural proposal cuts BEC iterations by 22%","U-Net speeds rotating condensate solver 14.5%","NN proposal cuts BEC density error by 46%","Neural net accelerates BEC solver, cuts iterations 22%","AI proposal trims rotating BEC iterations by 22%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000699,"raw_usage":{"total_tokens":3135,"prompt_tokens":901,"completion_tokens":2234,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":517,"completion_tokens_details":{"reasoning_tokens":2148}},"tokens_in":517,"tokens_out":2234,"duration_ms":17766,"temperature":1.0,"reasoning_tokens":2148,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:11:33.169861+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the rank correlation between the normalization error $e$ and the true density error $\\|\\tilde{\\rho} - \\rho^*\\|_{L^1}$ across the 40,000 training pairs; if the correlation is near zero, the acceptance gate is not selecting good proposals and the reported speedups would need a different explanation.","supporting_citations":[{"cited_title":"Peterseim, J","cited_arxiv_id":null,"evidence_quote":"Supplies the energy-adaptive Riemannian conjugate gradient method and its implementation that the neural network accelerates."},{"cited_title":"Ronneberger, P","cited_arxiv_id":null,"evidence_quote":"The U-Net convolutional architecture on which the accelerator is based."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The plane-wave density-functional toolkit used to discretize and run the Gross-Pitaevskii experiments."},{"cited_title":"Altmann, D","cited_arxiv_id":null,"evidence_quote":"Establishes the energy-adaptive Riemannian optimization framework on the Stiefel manifold that EARCG generalizes."},{"cited_title":"Henning and E","cited_arxiv_id":null,"evidence_quote":"Provides the contemporary overview of Gross-Pitaevskii numerical methods that motivates the benchmark and the comparison class."}],"review_version":1}