{"id":"af620c9e-37a8-462a-bce8-1ea753931e57","arxiv_id":"2509.14970","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A deep neural network trained on lattice QCD data provides the screening mass and coupling used to compute quarkonium dissociation temperatures, which roughly match earlier potential-model and lattice results.","lead":"A deep neural network trained on lattice QCD thermodynamics is used to set the screening mass and coupling in a quarkonium potential model, yielding dissociation temperatures for charmonia and bottomonia. The predicted melting temperatures roughly match the sequential suppression pattern seen in heavy-ion experiments.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The vacuum Cornell parameters α and σ are never specified, so the dissociation temperatures in Table I are not reproducible and the claimed lattice agreement cannot be audited.","rationale":"The reader's weakest_assumption identifies the same two linked issues: the unvalidated HTL conversion of quasi-parton masses to m_D/α_s, and the unspecified vacuum Cornell parameters. I agree that both are problematic. My stress-test singles out the missing α and σ as the most load-bearing because it is a direct internal gap: the paper's core numerical results cannot be reproduced or checked without these inputs. This is not a disagreement with the physics consensus; it is an under-determination within the paper's own framework. The reader's conditional verdict is appropriate. If the authors supply the parameters and the results are robust under reasonable variations, the concern would be resolved; until then, the quantitative agreement claim should remain conditional.","tokens_in":15349,"tokens_out":4126,"duration_ms":42248,"concrete_test":"Request the exact values of α and σ used for the Cornell potential from the authors (or from the code they say is available upon request), then recompute the dissociation temperatures in Table I with two alternative physical choices, e.g., (α = 0.2, σ = 0.20 GeV^2) and (α = 0.40, σ = 0.18 GeV^2), keeping all other inputs fixed. If the T_d values shift by more than ~0.1 T_c, or if an alternative choice moves any state outside its quoted bracket, then the unspecified potential parameters are load-bearing. As a secondary check, compare the ML m_D(T) from Fig. 2 with lattice determinations of the Debye mass from the color-singlet free energy over T/T_c ∈ [1,3]; a disagreement larger than 20% would further undermine the ML-input chain.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central quantitative claim is the set of dissociation temperatures in Table I. These are computed by solving the Schrödinger equation with the real part of a medium-modified Cornell potential (Eq. 4), which is built on the vacuum Cornell potential of Eq. (1). The parameters α and σ determine the strength of the Coulomb and string terms, and therefore directly control E_B(T), Γ(T), and both T_d criteria. Yet the paper never states the numerical values of α and σ used in the calculation. This is a load-bearing under-determination: the agreement with lattice QCD and with experiments such as CMS could be partly or wholly due to the choice of these two parameters, rather than to the ML-derived m_D(T) and α_s(T). The grid-convergence study in Appendix A.5 checks only numerical resolution, not parameter sensitivity. The DLQPM-to-m_D mapping in Eqs. (9)-(11) is also unvalidated—quasi-parton masses are fitted to the HotQCD equation of state, not to screening lengths, and the resulting m_D(T) in Fig. 2 is not compared to any independent lattice m_D determination—but the missing vacuum parameters are more directly threatening: even if the ML inputs were exactly right, the T_d predictions would still depend on the unspecified α and σ.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hybrid machine-learning/potential-model calculation of heavy-quarkonium dissociation temperatures in the QGP. A deep neural network (DLQPM) is trained on the HotQCD equation of state to extract quasi-parton masses; Eqs. (9)-(11) convert these masses into a running coupling α_s(T) and Debye screening mass m_D(T). These quantities enter a dielectric-screened Cornell potential (Eq. 4), whose real part is solved in the Schrödinger equation (Eq. 12) to obtain binding energies E_B(T), while the imaginary part yields the thermal width Γ(T) via Eq. (13). Dissociation temperatures are estimated with two criteria, Γ=2E_B and E_B=3T. Table I reports T_d ranges: Υ(1S) in [1.38,1.99] T_c, J/ψ in [1.13,1.30] T_c, Υ(2S) in [1.10,1.29] T_c, and ψ(2S) near or below T_c. The authors claim excellent agreement with lattice QCD results and interpret the hierarchy as consistent with sequential suppression observed in heavy-ion collisions.","tokens_in":15691,"tokens_out":5212,"duration_ms":57120,"significance":"The idea of using a lattice-constrained, ML-based quasi-particle description to feed a complex in-medium potential is a potentially useful bridge between equation-of-state data and quarkonium phenomenology. The paper is commendably explicit about the hybrid, perturbative-inspired character of the m_D and α_s extraction, and it includes numerical benchmarks and a grid-convergence appendix. However, the central quantitative results are currently not reproducible because the vacuum Cornell parameters and heavy-quark masses are never specified, and the mapping from DLQPM masses to m_D and α_s is not validated against independent determinations. If these issues are addressed, the framework could be a useful addition to the potential-model literature; in its present form the significance is conditional.","major_comments":[{"comment":"The central numbers in Table I depend directly on the vacuum Cornell parameters α and σ, which enter Eqs. (1) and (4), and on the heavy-quark masses m_c and m_b, which enter Eq. (12) through the reduced mass. None of these values is stated in the paper. The claim that the ML-enhanced potential 'tightens theoretical uncertainties' cannot be audited without knowing these inputs. The authors should state all parameter values, and ideally show how T_d varies under reasonable variations of α, σ, and m_Q. Without this, the agreement with lattice QCD and the sequential-suppression interpretation in Section III are under-determined.","section":"§II.A, Eq. (1), Eq. (4), Eq. (12), Table I"},{"comment":"The mapping from DLQPM quasi-parton masses to α_s(T) and m_D(T) is not an independent measurement. The quasi-parton masses are trained to reproduce the HotQCD equation of state, not screening lengths, and Eqs. (9)-(11) are leading-order HTL-inspired relations. Calling the resulting m_D and α_s 'non-perturbative, data-driven' overstates their status. The paper provides no comparison of the resulting m_D(T) with lattice determinations of the Debye mass, so the reader cannot assess whether this mapping is quantitatively reliable. The authors should benchmark m_D(T) against available lattice/effective-field-theory determinations and discuss the domain of validity of Eqs. (9)-(11).","section":"§II.B, Eqs. (9)-(11), Fig. 2"},{"comment":"Table IV shows that the dissociation temperatures continue to decrease monotonically as the grid is refined from N=4000 to N=6000 for every state and criterion (e.g., Υ(1S) with 2E_B=Γ goes from 1.992 to 1.956 T_c, and with E_B=3T from 1.375 to 1.349 T_c). The statement that the difference 'stabilizes for N≳3000' is not supported by the table; no N>6000 results or Richardson extrapolation are shown. Since Table I quotes the N=4000 values, the numerical convergence uncertainty should be quantified and propagated into the reported T_d ranges.","section":"Appendix A.5, Table IV"},{"comment":"There is an internal inconsistency in the numerical setup: §II.C states that the finite-difference solver uses Neumann boundary conditions, while Appendix A.1 and A.4 state that Dirichlet boundary conditions are imposed. This should be reconciled, and the effect of the boundary choice on weakly bound states near the dissociation point should be commented on.","section":"§II.C and Appendix A.1/A.4"}],"minor_comments":[{"comment":"Typographical slips: 'F uture Prospects' in Section IV.A and 'T raining set' in Section II.B should be corrected. Also, the notation for the number of flavors is inconsistent (n_f in Eq. (9), N_f in Eqs. (7) and (11)); the value of N_f (presumably 3) and N_c=3 should be stated explicitly.","section":"Headings and notation"},{"comment":"The authors state that they retrained the model ten times for uncertainty analysis, but no error bars are shown in Figs. 2-8 and Table I. The propagation of DLQPM uncertainties into m_D, α_s, E_B, Γ, and T_d should be quantified.","section":"Uncertainties"},{"comment":"The Data Availability section says code and data are available 'upon reasonable request.' For a paper whose main selling point is a data-driven ML framework, a public repository with the trained networks and the potential-model solver would greatly improve reproducibility.","section":"Code/data availability"},{"comment":"The Abstract and Section III claim 'excellent agreement with lattice QCD results,' but Table I compares mostly with other potential-model calculations and reviews. The authors should clarify which entries are direct lattice determinations and distinguish lattice QCD results from model-dependent extractions.","section":"Comparison with lattice QCD"}],"recommendation":"major_revision","confidential_remarks":"The core idea is interesting but the manuscript in its current form is not reproducible: the missing Cornell parameters and quark masses alone justify a major revision. The ML component is largely imported from earlier DLQPM papers (Refs. [35,36]); the novel quarkonium application is useful, but the authors should clearly delineate what is new here. I would encourage the editor to require a full parameter table, a validation of the m_D/α_s mapping against independent lattice data, and a proper convergence statement with N>6000 or Richardson extrapolation before reconsidering."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this is a straightforward application of the DLQPM model already published by the same group (refs 35,36) to the standard quarkonium dissociation calculation. The genuinely new output is the set of dissociation temperature brackets in Table I, e.g. Upsilon(1S) in [1.38,1.99] T_c. That is useful and plausibly correct in broad strokes.\n\nWhat the paper does well: the numerical solver is benchmarked properly against harmonic oscillator, Coulomb, and Woods-Saxon potentials; the dual-criterion approach (width vs binding and EB=3T) gives a physically reasonable bracket; and the authors are reasonably explicit that the DLQPM masses are fitted to the HotQCD equation of state and then converted to m_D and alpha_s via leading-order HTL formulas. That is a defensible hybrid, not a black box.\n\nNow the soft spots, in order of seriousness.\n\n1. The vacuum Cornell parameters alpha and sigma are never stated, and neither are the heavy quark masses. Equation (1) defines the potential, but the numbers used in the actual Schrödinger equation are absent. Since T_d depends directly on alpha and sigma, the central numbers are not reproducible and the claimed 'excellent agreement' with lattice cannot be audited. This is the load-bearing flaw.\n\n2. The grid convergence study in Appendix A5 shows the dissociation temperatures still drifting downward at N=6000. For Upsilon(1S) with the width criterion, T_d goes from 1.992 at N=4000 to 1.956 at N=6000, and the |delta| relative to N=500 is still growing. So the statement that results 'stabilize for N>=3000' is not supported; the main text uses N=4000, but the trend suggests the true value is a bit lower.\n\n3. The mapping from quasi-parton masses to m_D and alpha_s is not validated against any independent lattice determination of m_D. It's leading-order HTL structure with ML-fitted masses, so it's not an independent extraction. That by itself is fine, but the abstract's 'predicted' language overstates the status of m_D and alpha_s.\n\n4. The retraining uncertainty is quoted as mean and variance, but that uncertainty is not propagated to T_d. So the error bars in Table I—actually there aren't any—are missing.\n\nProportionate summary: the qualitative hierarchy is almost certainly right and the approach is worth engaging. But the missing parameters and the unresolved grid drift mean the numbers should not be taken at face value. This deserves peer review, but with the requirement that the authors specify alpha, sigma, m_c, m_b, propagate the ML uncertainty, and extend the convergence check to at least N=10000 or show that N=6000 is converged.\n\nFor a reading group, it's a maybe—it's a good case study of ML + phenomenology, but not a must-read. I wouldn't cite it in my own work until the parameters are stated.","headline":"A plausible and useful application of DLQPM to quarkonium melting, but missing potential parameters and unresolved grid drift make the numbers unreproducible as written.","tokens_in":16157,"tokens_out":3505,"would_cite":false,"duration_ms":36594,"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":"Deep neural networks trained on lattice QCD data yield a Debye mass and strong coupling that, inside a screened Cornell potential, reproduce the quarkonium melting hierarchy and sequential suppression seen in heavy-ion collisions.","keywords":["machine learning","deep neural network","quark-gluon plasma","heavy quarkonia","Debye screening mass","strong coupling","dissociation temperature","lattice QCD"],"falsifier":"Measure the Debye screening mass on the lattice, e.g., from the static quark-antiquark free energy, and compare with the DNN-extracted m_D(T) in Fig. 2 over the same T/T_c range; disagreement would falsify the HTL conversion that links quarkonium predictions to the lattice input. A second check is to repeat the calculation with the vacuum Cornell parameters alpha and sigma fixed to published values and see whether the reported T_d windows survive.","tokens_in":15269,"feed_emoji":"⚛️","tokens_out":12261,"duration_ms":111541,"temperature":0.7,"pith_summary":"The paper tries to establish that machine-learned quasi-parton masses, extracted from lattice QCD equation-of-state data by a deep residual network, can supply the two temperature-dependent inputs — Debye screening mass m_D(T) and strong coupling alpha_s(T) — needed for a complex screened Cornell potential. From that potential, the radial Schrödinger equation is solved for J/psi, psi(2S), Upsilon(1S), and Upsilon(2S), giving binding energies and thermal widths. Dissociation temperatures are read off with two criteria, a width condition (Gamma = 2E_B) and a thermal-energy condition (E_B = 3T). The resulting windows — Upsilon(1S) at 1.38-1.99 T_c, J/psi at 1.13-1.30 T_c, Upsilon(2S) at 1.10-1.29 T_c, psi(2S) at or below T_c — reproduce the sequential suppression hierarchy observed in heavy-ion collision experiments. If the claim is right, ML-augmented potentials become a reliable bridge from lattice QCD to quarkonium phenomenology.","feed_headline":"Deep learning gives quarkonia melting windows from lattice QCD data","feed_subtitle":"A DNN trained on the QCD equation of state reproduces the sequential suppression of charmonia and bottomonia.","key_machinery":"The paper's central object is the Deep-Learning Quasi-Parton gas Model (DLQPM), an eight-layer residual neural network with 32 neurons per layer that maps temperature to quasi-parton masses for gluons, light quarks, and strange quarks. The network is trained by minimizing the mean-squared error of the entropy density and trace anomaly against lattice QCD data over T/T_c in [1,3], with an additional high-temperature constraint nudging the gluon/light-quark mass ratio toward the hard-thermal-loop value. The learned masses are then inserted into Eq. (10) to define alpha_s(T) and Eq. (11) to define the Debye mass m_D(T). These two quantities set the real and imaginary parts of the complex Cornel","core_discovery":"The central result is a data-driven replacement for perturbative medium parameters: a deep residual network is trained to match lattice QCD entropy density and trace anomaly, and the learned quasi-parton masses are converted by leading-order hard-thermal-loop relations into m_D(T) and alpha_s(T). These enter a complex Cornell potential (real part screens, imaginary part encodes Landau damping). Solving the Schrödinger equation yields, per state, the temperature where thermal width equals twice the binding energy and where binding energy equals 3T. The reported T_d values for J/psi, psi(2S), Upsilon(1S), Upsilon(2S) form a dissociation window that the paper interprets as agreeing with lattice","pith_inferences":["The same learned quasi-parton masses could supply m_D(T) and alpha_s(T) to other medium-sensitive calculations — parton energy loss, thermal photon or dilepton production, transport coefficients — giving one lattice-anchored set of inputs across QGP phenomenology.","A sharper test of the method would train the network directly on lattice data for the heavy-quark potential (from Wilson loops or spectral functions), letting the DNN learn the full complex potential and bypass the analytic Cornell-plus-dielectric form.","The width and binding criteria bracket T_d, but a direct extraction of the spectral function peak at each temperature from lattice QCD could pin the melting temperature to a single value and adjudicate between the two brackets."],"forward_implications":["Upsilon(1S) would survive to at least 1.38 T_c and possibly up to 1.99 T_c, supporting its use as a probe of the hottest deconfined phase.","J/psi melting in the 1.13-1.30 T_c window leaves room for both a surviving primordial component and late-stage recombination, matching the measured incomplete suppression.","psi(2S) melting at or below T_c explains why the excited charmonium state is dramatically more suppressed than the ground state.","The dual-criterion brackets mean that comparisons with lattice spectral functions should be made against a temperature band rather than a single number.","The lattice-anchored m_D(T) and alpha_s(T) reduce the reliance of quarkonium calculations on purely perturbative inputs."],"fun_headline_variants":["Neural nets map quarkonia meltdown in QGP","DNN predicts charmonia and bottomonia dissociation","Deep learning pinpoints quarkonia melting windows","Lattice-trained DNN yields quarkonia suppression","Machine learning decodes quarkonium meltdown"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that quasi-parton masses fit only to the equation of state can be converted, via leading-order hard-thermal-loop formulas, into the exact Debye mass and running coupling that control a complex Cornell potential — with the vacuum Cornell parameters alpha and sigma not specified in the paper; if that conversion is not valid, the dissociation temperatures do not follow.","fun_headline_variants_meta":{"raw":{"variants":["Neural nets map quarkonia meltdown in QGP","DNN predicts charmonia and bottomonia dissociation","Deep learning pinpoints quarkonia melting windows","Lattice-trained DNN yields quarkonia suppression","Machine learning decodes quarkonium meltdown"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000531,"raw_usage":{"total_tokens":2428,"prompt_tokens":813,"completion_tokens":1615,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":557,"completion_tokens_details":{"reasoning_tokens":1543}},"tokens_in":557,"tokens_out":1615,"duration_ms":14201,"temperature":1.0,"reasoning_tokens":1543,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T16:13:02.992914+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the Debye screening mass on the lattice, e.g., from the static quark-antiquark free energy, and compare with the DNN-extracted m_D(T) in Fig. 2 over the same T/T_c range; disagreement would falsify the HTL conversion that links quarkonium predictions to the lattice input. A second check is to repeat the calculation with the vacuum Cornell parameters alpha and sigma fixed to published values and see whether the reported T_d windows survive.","supporting_citations":[],"review_version":1}