{"id":"dac9d787-9584-4bab-ab48-562a301a515a","arxiv_id":"2506.10882","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A DeePMD-based machine learning potential reproduces CO ice desorption dynamics with 11,000 trajectories and reveals a near-linear rise of rotational excitation with translational energy.","lead":"Researchers built a machine-learning model of solid carbon monoxide that can simulate how a vibrationally excited CO molecule kicks another molecule off an ice cluster. Running 11,000 simulations instead of 100, they mapped how desorbed molecules split energy between motion and rotation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The MLP is validated only on its own training trajectories; the 11,000-MD ensemble may sample out-of-distribution configurations, and the 77% vs 88% desorption-yield difference is a borderline symptom of possible systematic bias.","rationale":"The reader's weakest assumption is exactly the one I identified: the MLP is trained and validated on the same 100 AIMD trajectories, and there is no test of its accuracy on the 11,000-trajectory ensemble. This is load-bearing because the paper's headline physical result—the detailed translational/rotational distributions and the new Jpeak/Jmax correlation—is obtained exclusively from the MLP trajectories. If the MLP extrapolates poorly outside the training distribution, those distributions would be biased even though they are smooth. The desorption-yield difference (77% vs 88%) provides a concrete, if preliminary, indication that the MLP may not exactly reproduce the AIMD dynamics. The proposed DFT re-evaluation of sampled MLP-trajectory configurations is a direct and computationally feasible check: it requires only single-point DFT calculations on configurations the MLP itself produced, and it would settle whether the MLP's dynamics stayed within the region where it is accurate. I agree with the reader that the paper warrants conditional acceptance rather than full acceptance, because the missing out-of-distribution validation is the key uncertainty. No ad hominem is intended; the concern is purely about the evidence for the central claim. If the proposed test passes, the conditional status could be upgraded; if it fails, the MLP-based conclusions would need to be revised.","tokens_in":13421,"tokens_out":10482,"duration_ms":124041,"concrete_test":"Uniformly sample 1000–2000 configurations from the 11,000 MLP trajectories, spanning all simulation times (including t > 2 ps and desorbed molecules at distances > 3 Å from the aggregate surface), and recompute their energies and forces with VASP at the same PBE+D3(BJ) level used in training. Compare MLP predictions to DFT: if the energy RMSE on these out-of-training configurations is comparable to the validation RMSE (75 meV) and the force RMSE stays below ~0.3 eV/Å for all atom types, the generalization assumption holds and the 11,000-trajectory results are reliable. If the RMSE is significantly larger (e.g., energy error > 150 meV) or the errors concentrate on desorbed molecules or highly excited ground-state CO bonds, then the MLP's energy distributions and the Jpeak/Jmax linear trend are not trustworthy, and the conditional acceptance should be reconsidered.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the MLP accurately describes the desorption process and its energy distributions. The evidence for this is the energy/force RMSE on a validation set, the dissociation-curve agreement, and the match to AIMD/experiment. However, every validation check in Section III.A is performed on configurations drawn from the same 100 AIMD trajectories used for training (Section II.A). The 11,000 LAMMPS trajectories (Section II.B) start from 11,000 independent thermalized configurations and evolve for 5 ps, generating configurations that are not part of the training set. Nothing in the paper verifies that the MLP remains accurate on these out-of-training configurations, including the late-time desorbed molecule at large distances from the aggregate, or on rare high-energy events such as a ground-state CO molecule being vibrationally excited beyond the v≈14 region where the C/O network was tested (Fig. 6). The desorption yield from the MLP (77%) differs from the AIMD benchmark (88%) by 11 percentage points; given only 100 AIMD trajectories, this is a borderline but not negligible discrepancy. If the MLP has a small systematic bias, the improved statistics of 11,000 trajectories could produce energy distributions that are smooth but shifted, and the new Jpeak/Jmax linear trend (Fig. 10) would inherit that bias. Thus the load-bearing assumption is that the MLP generalizes from the training trajectories to the full ensemble used for the physics conclusions.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a deep learning-based machine learning potential (MLP) for solid CO, constructed with DeePMD-kit, in which the vibrationally excited CO molecule is treated by two distinct atom types (X1 and X2). The MLP is trained on approximately 18,000 configurations drawn from 100 AIMD trajectories of vibrational relaxation from v=40 in a 50-molecule CO aggregate. Validation shows low energy RMSE (0.075 eV) and moderate force RMSE for the X1/X2 atoms, plus good agreement with a DFT dissociation curve. The authors then run 11,000 LAMMPS MD trajectories with the MLP, reporting a 77% desorption yield (compared to 88% in the prior AIMD study), smooth translational and rotational energy distributions that agree with experimental data, and a new result: a near-linear increase of J_peak and J_max with translational energy. The paper claims the MLP accurately describes the desorption process and provides a significant improvement in statistical sampling over AIMD.","tokens_in":13799,"tokens_out":6947,"duration_ms":75364,"significance":"If the validation concerns are resolved, this work offers a valuable methodological contribution: the selective X1/X2 treatment of a highly excited molecule is a promising strategy for extending MLPs to non-equilibrium, strongly anharmonic dynamics in condensed phases. The 11,000-trajectory ensemble provides a substantial gain in statistical power, enabling distribution comparisons with experiment that were previously impractical with AIMD. The agreement of the MLP rotational and translational distributions with experimental data is encouraging and supports the underlying desorption mechanism. However, the central claim of accuracy is currently weakened by the in-sample validation and by an unexplained 11 percentage-point discrepancy in the desorption yield relative to AIMD. The new linear-correlation finding also rests on only four points without statistical treatment for J_max. These issues need to be addressed before the results can be considered fully supported.","major_comments":[{"comment":"The MLP desorption yield (77%) differs from the AIMD yield (88%) by 11 percentage points. With 100 AIMD trajectories, the standard error of the AIMD yield is about 3.3% (sqrt(0.88*0.12/100)), and with 11,000 MLP trajectories the MLP yield has a standard error below 0.5%, so the difference is roughly 3 standard deviations. The text describes this as 'close' but provides no statistical comparison or discussion. Because the desorption yield is a primary observable of the process, this discrepancy may indicate a systematic bias in the MLP, a difference in initial-condition sampling, or a real physical difference. Please either add confidence intervals and a formal comparison, or run AIMD on a subset of the same initial conditions used in the MLP trajectories to discriminate between these possibilities.","section":"III.B first paragraph"},{"comment":"All validation checks—energy parity (Fig. 2), force parity (Figs. 3-5), and the dissociation curve (Fig. 6)—are performed on configurations drawn from the same 100 AIMD trajectories that supplied the training set. The 11,000 production trajectories start from a new thermalization ensemble and evolve over 5 ps, generating configurations (including desorbed molecules far from the aggregate) that are not part of the validation set. The paper does not demonstrate that the MLP remains accurate on these out-of-training configurations. Please validate the MLP on a hold-out set of AIMD trajectories not used in training, or re-evaluate a sample of MLP-generated configurations with DFT and compare energies and forces, or at minimum discuss the in-sample nature of the validation and its implications for the reported uncertainties on desorption yields and energy distributions.","section":"III.A and II.A"},{"comment":"The claim that J_peak and J_max increase 'nearly linearly' with translational energy is based on only four data points for each quantity. J_max is defined by a 3%-of-maximum threshold that is sensitive to sampling noise, yet no uncertainty is given for J_max; the error bars shown for J_peak are not tied to a regression. Without a linear fit with confidence intervals, or bootstrap estimates for J_max, the linear trend is not quantitatively supported. Please provide uncertainties for J_max and perform a proper regression (including R² and parameter uncertainties), or soften the claim to a qualitative observation.","section":"III.C, Figure 10"},{"comment":"The X1/X2 force RMSE values (0.186 and 0.154 eV/Å) are roughly four times larger than the C/O values (0.045 and 0.038 eV/Å). The authors argue that this is acceptable because the absolute forces on X1/X2 are larger, but the excited molecule drives the entire energy-transfer process, so a systematic force error on these atoms could directly bias the desorption yield and the final energy distributions. The paper does not quantify how these force errors propagate into the observables. Please add a sensitivity test—for example, running a few MLP trajectories with a different MLP seed, or comparing a handful of MLP trajectories against AIMD from the same initial conditions—to estimate the impact of X1/X2 force errors on the reported results.","section":"III.A, Figures 3-4"}],"minor_comments":[{"comment":"There are several typographical errors: 'pannel' should be 'panel' in the captions of Figs. 8, 9, and 10; 'corelation' should be 'correlation' in Section III.A; 'mentionned' should be 'mentioned' in Section II.A; and 'the later' should be 'the latter' in the same section.","section":"Throughout"},{"comment":"The dataset sampling intervals overlap: the first interval is 'until t=500 fs', the second is 'between 100 fs and 2500 fs', and the third is 'above 2500 fs'. Clarify the boundary times (e.g., whether 100 fs and 2500 fs belong to two intervals) to avoid ambiguity.","section":"II.A"},{"comment":"The colorbar label 'Absolute Error (eV)' appears truncated in the displayed figure; please check that all axis and colorbar labels are fully visible.","section":"Figure 2"},{"comment":"The phrase 'High values distributions are computed for translational energies lower than 200 meV' is unclear; please rephrase to specify the energy window used for the high-translational-energy panel.","section":"III.C, Figure 8 caption"},{"comment":"The procedure for setting the internal momentum to v=40 is not described in detail (e.g., whether the bond length and velocities are set deterministically or sampled from a distribution). Please specify the exact method for reproducibility.","section":"II.B"},{"comment":"The paper does not state whether the trained MLP weights or the trajectory data will be made available. If the journal's policy requires data/code availability, please add a statement.","section":"Data availability"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of the journal. The X1/X2 treatment of the excited molecule is an interesting methodological idea, and the large statistical ensemble is a clear strength. However, the validation is in-sample, and the desorption yield discrepancy between the MLP and AIMD is larger than expected from statistical noise. These are load-bearing issues for the paper's central claim of accuracy. The linear-correlation claim also needs stronger statistical support. I recommend major revision, but I believe the issues are fixable with additional validation and a more careful statistical treatment."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis is a genuine step forward for MLP-based simulation of vibrationally excited molecules in condensed phases. The X1/X2 idea—separate neural networks for the two atoms of the excited molecule—is simple, effective, and reusable. The leap from 100 AIMD trajectories to 11,000 MLP trajectories gives smooth translational and rotational distributions that match experiment better than before, and the resolved Jpeak/Jmax versus translational energy trend is new, even if it only spans four bins.\n\nI'd send this to peer review. But there are soft spots you should know about.\n\nThe validation is entirely in-sample. Every RMSE, the force-direction heatmaps, and the dissociation curve are computed on configurations drawn from the same 100 AIMD trajectories used for training. The 11,000 production runs generate configurations—especially late-time desorbed molecules far from the aggregate—that are never checked against DFT. That is a legitimate gap, and the natural fix (re-evaluate a few hundred LAMMPS snapshots with DFT) is cheap and obvious.\n\nSecond, the desorption yield difference is understated. The MLP gives 77%, AIMD 88%. With 100 AIMD trajectories, the counting error on 88% is about ±3%, so an 11-point gap is roughly three sigma, not 'close.' The paper does not discuss this. It could reflect a small systematic bias in the MLP, which is exactly the kind of thing that could shift the energy distributions in the 11,000-run ensemble.\n\nThird, the linear Jpeak/Jmax trend rests on four points. Jpeak has 95% error bars, Jmax does not, and the '3% of maximum' cutoff for Jmax is arbitrary. The trend looks real, but the evidence is thinner than the abstract implies.\n\nNone of this is disqualifying. The match with experimental translational and rotational distributions is a genuine out-of-sample check of sorts, even if the experimental comparison is qualitative. The authors are transparent about the training data limits. The method and the physics are both worth the time of a serious referee. I'd ask for the out-of-sample validation, a candid discussion of the yield gap, and ideally the training data and MLP files to be deposited. With those changes, the paper is a solid contribution.\n\nFor a reading group: yes—it's a good case study in where MLP validation can miss.","headline":"Solid, reusable MLP study with a clever excited-molecule trick; the validation gap and yield mismatch matter, but the paper deserves review.","tokens_in":14289,"tokens_out":2936,"would_cite":true,"duration_ms":34713,"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":"A deep learning potential trained on ab initio data reproduces the desorption of CO from an ice aggregate and, with 11,000 trajectories, exposes a nearly linear rotational-translational energy coupling.","keywords":["machine learning potential","CO ice","photodesorption","vibrational relaxation","desorption dynamics","interstellar ices","deep potential","translational-rotational coupling"],"falsifier":"Run a set of new AIMD trajectories from the same 11,000 initial conditions (or from a large random subset) and compare the desorption yield, desorption times, and the $J_{\\mathrm{peak}}$ versus translational-energy trend with the MLP predictions; a statistically significant disagreement in any of these observables would falsify the claim that the potential generalizes. In addition, one could compute DFT forces directly for configurations sampled from the MLP trajectories in the late desorption phase, where the escaping molecule is far from the aggregate, and look for systematic force errors that would bias the desorption outcome.","tokens_in":13272,"feed_emoji":"🧊","tokens_out":11872,"duration_ms":117605,"temperature":0.7,"pith_summary":"The paper claims that a machine-learned interatomic potential, trained on data from 100 ab initio molecular dynamics trajectories, can reproduce the full desorption of a carbon monoxide molecule from an aggregate of 50 CO molecules after vibrational excitation to $v=40$, matching prior theory and experiment. The key trick is to give the two atoms of the vibrationally excited CO their own dedicated neural networks, separate from the 49 ground-state molecules, so that the potential can represent the strongly anharmonic high-energy stretch without destabilizing the dynamics. With this potential the authors run 11,000 independent trajectories, more than 100 times the previous ab initio sample, and obtain smooth translational and rotational energy distributions of the desorbed molecules. The finer statistics reveal that both the most probable and the maximum rotational quantum number, $J_{\\mathrm{peak}}$ and $J_{\\mathrm{max}}$, rise nearly linearly with translational energy, evidence of coupled translational–rotational energy flow during desorption. If the claim holds, it makes systematically sampled simulations of interstellar ice photodesorption affordable.","feed_headline":"Neural network reproduces CO ice desorption with 11,000 trajectories","feed_subtitle":"A trained potential matches ab initio and experiment, revealing how rotation scales with translational energy.","key_machinery":"The central object is the machine-learned interatomic potential used to run the production dynamics. It is an atom-centered high-dimensional neural network potential in which the total energy is the sum of atomic energy contributions, each computed by a descriptor and a fitting network; the paper's innovation is to give the two atoms of the vibrationally excited CO molecule separate networks, X1 and X2, so that the strongly anharmonic potential-energy region of the stretched bond is fitted independently of the ground-state molecules. This separation is what allows the potential to stay accurate up to the $v=40$ energy (8.26 eV) for the hot molecule while the surrounding molecules are described by the standard carbon and oxygen networks. The potential reproduces the DFT dimer dissociation curve up to 8.26 eV for X1–X2 and up to about 6.5 eV for the ordinary CO networks, and the production trajectories are started from the same kind of thermalized aggregate ensemble and integrated with the same 0.1 fs time step as the prior AIMD study, making the statistical comparison direct.","core_discovery":"On its own terms, the paper establishes that a neural network potential of the atom-centered type can describe, at near-DFT accuracy, the relaxation and desorption dynamics that follow depositing 8.26 eV of vibrational energy into one CO molecule inside a 50-molecule CO ice aggregate. The authors train the potential on configurations sampled from 100 DFT-based AIMD trajectories, and they assign separate neural networks, labelled X1 and X2, to the carbon and oxygen atoms of the excited molecule so that the short, strongly anharmonic C–O bond of the hot molecule is not handled by the same parameters as the bonds of the cold surroundings. In 11,000 microcanonical trajectories started from a 15 K thermal ensemble, 77% of excitations lead to desorption, close to the 88% found in AIMD, with an average desorption time of about 2 ps and no desorption before 600 fs. The desorbed CO is predominantly $v=0$ (about 5% in $v=1$), and the translational distribution peaks near 25 meV and falls to zero above about 800 meV, reproducing the experimental time-of-flight data. With the improved sampling, rotational distributions can be resolved in narrow translational windows, and they show $J_{\\mathrm{peak}}$ increasing from 4 to 8 and $J_{\\mathrm{max}}$ from 25 to 44 as the translational energy interval moves from 0–50 meV to 150–200 meV; the paper interprets this near-linear rise as continuous, coupled rotational–translational energy redistribution during desorption.","pith_inferences":["If the potential generalizes beyond the $v=40$ pathway it was trained on, the same architecture could scan other initial vibrational levels, ice temperatures, and aggregate sizes at essentially no additional ab initio cost; the paper does not demonstrate this.","The linear rotational-translational correlation found for CO may be a generic property of collision-mediated desorption from weakly bound molecular ices; testing on N2, O2, or CO2 ices would indicate whether the mechanism is universal.","The X1/X2 separation is effectively a 'tagged molecule' trick; it could be reused to follow energy flow out of any selected excited species in a multichromophore system, not just CO.","The 77% versus 88% desorption yield difference may be within statistical scatter, but it could also mark a small bias in the potential's rare-event regions; an independent MLP retrained from different random weights, or more AIMD trajectories, would settle this."],"forward_implications":["The roughly 1500-fold lower cost per energy-and-force evaluation makes statistically converged trajectory ensembles for desorption studies feasible, something AIMD cannot deliver.","The method can be transferred to other localized excitations in condensed-phase molecular systems, as the X1/X2 separation is general.","The resolved rotational distributions provide a quantitative map of how vibrational relaxation energy is partitioned into translation and rotation of the desorbing molecule.","The near-linear $J_{\\mathrm{peak}}$ and $J_{\\mathrm{max}}$ versus translational energy relationship offers a testable signature of the indirect desorption mechanism in future experiments."],"supporting_citations":[{"why":"Supplies the 100 AIMD trajectories used to build the training set and the ab initio desorption yield and energy distributions used as the primary theoretical comparison.","marker":"[25]"},{"why":"Provides the experimental rotational and translational energy distributions of desorbed CO used to validate the MLP results.","marker":"[26]"},{"why":"Provides the deep-learning interatomic potential training package used to construct and fit the MLP.","marker":"[33]"},{"why":"Documents the descriptor and training procedure of the deep potential toolkit, on which the MLP architecture relies.","marker":"[34]"},{"why":"Establishes the atom-centered neural network potential formalism that underlies the MLP's energy decomposition.","marker":"[35]"},{"why":"Provides the molecular dynamics engine used to run the 11,000 production trajectories with the MLP.","marker":"[45]"}],"fun_headline_variants":["Neural net mimics CO ice desorption in 11,000 runs","Machine learning potential nails CO ice desorption dynamics","Deep learning reveals rotational-translational coupling in CO desorption","CO ice desorption: neural network matches experiment at 11k trajectories"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The neural network potential, trained on configurations sampled from 100 AIMD trajectories of the same desorption process, generalizes accurately to the full ensemble of 11,000 independent initial conditions, including the regions where the excited molecule and the desorbing molecule move far from the aggregate.","fun_headline_variants_meta":{"raw":{"variants":["Neural net mimics CO ice desorption in 11,000 runs","Machine learning potential nails CO ice desorption dynamics","Deep learning reveals rotational-translational coupling in CO desorption","CO ice desorption: neural network matches experiment at 11k trajectories"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000502,"raw_usage":{"total_tokens":2535,"prompt_tokens":1108,"completion_tokens":1427,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":724,"completion_tokens_details":{"reasoning_tokens":1355}},"tokens_in":724,"tokens_out":1427,"duration_ms":12503,"temperature":1.0,"reasoning_tokens":1355,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:15:26.405712+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a set of new AIMD trajectories from the same 11,000 initial conditions (or from a large random subset) and compare the desorption yield, desorption times, and the $J_{\\mathrm{peak}}$ versus translational-energy trend with the MLP predictions; a statistically significant disagreement in any of these observables would falsify the claim that the potential generalizes. In addition, one could compute DFT forces directly for configurations sampled from the MLP trajectories in the late desorption phase, where the escaping molecule is far from the aggregate, and look for systematic force errors that would bias the desorption outcome.","supporting_citations":[{"cited_title":"Del Fr \\'e , author A","cited_arxiv_id":null,"evidence_quote":"Supplies the 100 AIMD trajectories used to build the training set and the ab initio desorption yield and energy distributions used as the primary theoretical comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the experimental rotational and translational energy distributions of desorbed CO used to validate the MLP results."}],"review_version":1}