Pith. sign in

REVIEW 3 major objections 5 minor 5 references

XMCQDPT2-Fidelity Transfer-Learning Potentials and a Wavepacket Oscillation Model with Power-Law Decay for Ultrafast Photodynamics

T0 review · 3 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The paper claims that transfer-learned neural-network potentials fine-tuned on roughly 1,450 XMCQDPT2 points can drive nonadiabatic dynamics that expose a slow ~200 fs photodissociation channel in the methaniminium cation.

desk verdict Solid transfer-learned MLIP work for CH2NH2+ at XMCQDPT2 level, but the kinetic 'power-law' model is a fitted exponential in disguise and the claimed independent validation of the sigma-pi*/S0 pathway does not hold up. read the letter →

arxiv 2512.07537 v1 pith:X3PX4UNL submitted 2025-12-08 physics.chem-ph

Ivan V. Dudakov , Pavel M. Radzikovitsky , Dmitry S. Popov , Denis A. Firsov , Vadim V. Korolev , Daniil N. Chistikov , Vladimir E. Bochenkov , Anastasia V. Bochenkova This is my paper · ORCID
classification physics.chem-ph
keywords machinelearninginteratomicpotentialsnonadiabaticdynamicsconicalintersectionstransferXMCQDPT2Landau-Zenersurfacehoppingmethaniminiumcationpower-lawkinetics
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to show that machine-learning potentials trained on a deliberately small set of high-level quantum-chemistry data can reproduce the full excited-state decay of the methaniminium cation, a molecule that models protonated Schiff bases and is relevant to Titan's atmosphere. The authors fine-tune an equivariant neural-network potential from a cheaper CASSCF description to XMCQDPT2 accuracy using about 1,450 reference points, then run 600-trajectory Landau-Zener surface-hopping dynamics from the S2 state. They also introduce a wavepacket oscillation model in which repeated attempts to cross a conical intersection produce power-law decay, letting them extract state-specific lifetimes rather than fit exponentials. Fits to the simulated populations require two distinct S1-to-S0 routes, with lifetimes of 25 fs and ~200 fs; the slow route supports a recently proposed photochemical pathway through a sigma-pi*/S0 conical intersection. If the approach holds, high-level nonadiabatic dynamics becomes affordable for other photochemical systems.

What carries the argument

The load-bearing object is the multi-passage survival cascade model: a wavepacket oscillates near a conical intersection with period T, crossing the intersection twice per cycle, and each crossing succeeds with Landau-Zener probability P_LZ; the chance of surviving N attempts is (1-P_LZ)^N, which produces power-law kinetics and a mean lifetime T/(2P_LZ). This converts a single quantum transition probability into a classical rate constant without assuming exponential decay. On the machine-learning side, the enabling mechanism is transfer learning: a cheaper pretrained model supplies the representation, and a small fine-tuning set transfers it to XMCQDPT2 fidelity; an ensemble's energy-standar

What would settle it

Run the same S2-initiated dynamics with a surface-hopping method that uses explicit nonadiabatic coupling vectors at the XMCQDPT2 level and compare mean hopping times and S1 branching; if the 28.4 fs S1 time or the 84:16 split changes substantially, the Landau-Zener approximation is at fault. Alternatively, an ultrafast pump-probe experiment on CH2NH2+ that resolves a ~200 fs component in the S1 population would confirm the minor sigma-pi*/S0 channel.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a single-state transfer-learning protocol—pretrain an equivariant neural network on roughly 50,000 CASSCF geometries, then fine-tune on ~1,450 XMCQDPT2 points—yields potential energy surfaces that reproduce XMCQDPT2 reference conical intersections to within ~0.2 eV in energy and ~0.1 Å in geometry. These potentials drive Landau-Zener surface-hopping dynamics that map all decay channels after S2 excitation, giving uncertainty-corrected mean hopping times of 5.0 fs for S2→S1 and 28.4 fs for S1→S0, and a product distribution dominated by CN cleavage with ~2% direct H2 loss. Fitting the populations with the new wavepacket oscillation model

Load-bearing premise

The entire dynamical picture rests on Landau-Zener single-passage probabilities—computed without nonadiabatic coupling vectors—fully determining whether a trajectory hops, and the kinetic model then assumes those probabilities, the oscillation periods, and the independence of repeated passages all stay constant.

Editorial extensions

If this is right

  • High-level multi-reference nonadiabatic dynamics can be run on the fly with roughly 1,450 XMCQDPT2 reference points, as long as a cheaper pretrained potential is available for transfer learning.
  • Ensemble uncertainty weighting changes real conclusions: it shortens the S2→S1 mean hopping time from 14.6 fs to 5.0 fs and brings models trained differently into agreement, so it should be a standard part of ML-driven dynamics.
  • Exponential decay is not the right default for repeated conical-intersection passages; lifetimes extracted from single-exponential fits to ultrafast populations need to be revisited under the multi-passage picture.
  • The S1 population decay cannot be fit without splitting it into two routes; the minor sigma-pi*/S0 route has a ~200 fs lifetime and 16% branching, supporting the recently discovered photochemical pathway.
  • Branching ratios among photodissociation products are robust to the choice of MLIP, with CN cleavage dominant and direct H2 loss at ~2% within 100 fs.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the transfer-learning recipe generalizes, the same ~1,500-point fine-tuning could bring XMCQDPT2-fidelity dynamics to larger protonated Schiff bases and other chromophores where cheaper CASSCF datasets already exist.
  • A concrete testable prediction follows from the 16% branching but only 2% yield at 100 fs: trajectories longer than 100 fs should show delayed H2-loss products appearing on the ~200 fs timescale.
  • The model's lifetime formula T/(2P_LZ) suggests that experimental pump-probe transients could be fit directly with power laws to estimate single-passage Landau-Zener probabilities and oscillation periods, rather than phenomenological exponentials.
  • The mismatch between the 13.6 fs fitted S2 lifetime and the 5.0 fs trajectory-averaged hopping time hints that the constant-period, constant-probability assumption breaks down near the Franck-Condon region; a time-dependent version of the cascade model would reconcile the two.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper develops machine-learned interatomic potentials (MLIPs) targeting XMCQDPT2/SA(3)-CASSCF(12,12) accuracy for the lowest three singlet states of the methaniminium cation, using transfer learning, Δ-learning, and multi-state architectures. These potentials are used in Landau–Zener surface-hopping (LZBL) dynamics initiated from S2, and an ensemble uncertainty weighting is introduced to correct state populations. The authors report that transfer learning improves test-set errors, that MECI geometries are reproduced within ~0.1 Å, and that uncertainty corrections bring different MLIP models into closer agreement. They also introduce a 'wavepacket oscillation model' with parameters P_LZ and T, fit to the simulated population traces, from which channel-specific lifetimes are extracted. The paper claims that these kinetic fits require branching in S1 and thereby support the existence of a σπ*/S0 conical-intersection pathway for direct H2 loss.

Significance. If the MLIP methodology and the kinetic interpretation were both fully validated, the work would be a valuable contribution: it demonstrates a practical route to XMCQDPT2-fidelity excited-state dynamics with a large active space, and it proposes a transparent bridge between nonadiabatic surface hopping and classical kinetics. The transfer-learning gains (Table I) and MECI validation (Table II) are concrete and support the MLIP component. The uncertainty-weighting procedure (Eq. 1) is a useful diagnostic and the observation that it reconciles models is interesting. However, the kinetic model at the center of the paper's headline claims is internally inconsistent and its evidential basis for the σπ*/S0 pathway is not statistically established. These issues are load-bearing for the concluding claims about channel-specific lifetimes and pathway validation.

major comments (3)
  1. [Sec. III.D, Eqs. (2)–(9)] The lifetime extraction is circular. P_LZ and T are fitted to the simulated population curves in Fig. 4, and Eq. (9) then defines τ = T/(2P_LZ). Reporting these as 'extracted lifetimes' is a restatement of fit parameters, not an independent prediction. The only model-free observables are the average hopping times in Table III; the paper itself notes discrepancies between these and the fitted lifetimes. To support the claimed predictive character, the authors should either fit P_LZ and T from ab initio Landau–Zener parameters and oscillation frequencies and then compare with the simulated populations, or clearly label the results as a descriptive fit and provide goodness-of-fit statistics and parameter uncertainties.
  2. [Sec. III.D, Eqs. (2)–(7)] The 'power-law' label contradicts the model equations. With constant P_LZ and N(t) linear in t (Eqs. 6–7), (1−P_LZ)^{N(t)} = exp[N(t) ln(1−P_LZ)] is an exponential decay, not a power law. Furthermore, the geometric distribution with constant P_LZ is memoryless; the statement that 'each failed hopping attempt increases the probability of successful hopping' is inconsistent with the model's own assumption of independent passages. The terminology appears throughout the abstract and conclusions. The model should be renamed, or its equations modified (e.g., with a passage-dependent P_LZ) to genuinely produce nonexponential decay.
  3. [Sec. III.D, Table IV] The claim that a high-quality fit requires explicit S1 branching is not supported by any quantitative model comparison. No χ², AIC/BIC, residual analysis, or cross-validation is provided for a one-channel (α=1) versus two-channel (Eq. 3) model. Since the two-channel model adds three parameters (P_LZ^b, T_b, α), it will generally fit better even if the second channel is spurious. Moreover, the minor-channel lifetime of 222 fs greatly exceeds the 100 fs simulation window, and the 16% branch fraction is a fitted extrapolation, not an observed product yield (direct H2 loss is 2% at 100 fs). Please report a formal comparison against the single-channel model and, ideally, identify the σπ*/S0 pathway directly from trajectory geometries rather than from a fitted branch.
minor comments (5)
  1. [Sec. III.D, text near Table IV] The text states 'Landau-Zener probability P_LZ^(b)=0.60' for the dominant channel; this should be P_LZ^(a)=0.60, consistent with Table IV.
  2. [Methods II.G heading] Typo: 'photodissotiation' should be 'photodissociation'; also group 7 contains 'disostiation'.
  3. [Fig. S5/S6 reference] Grammar: 'Figures S5 and S6 shows' should be 'show'.
  4. [Data Availability] The statement 'available from the corresponding author upon reasonable request' is less reproducible than depositing the trained models and simulation data in a public repository; consider sharing at least the key training sets and fitted kinetic parameters.
  5. [References] Ref. 38 is a companion arXiv preprint; if it is under review, this should be stated explicitly. The claim that metrics are 'competitive with X-MACE' would benefit from a direct table or figure.

Circularity Check

3 steps flagged · score 6.0 of 10

Kinetic lifetimes are closed-form restatements of fitted P_LZ and T; the claimed support for the sigma-pi*/S0 pathway is built into Eq. (3) and then confirmed by fitting its own branch weight, with no single-channel comparison.

  1. fitted input called prediction [Section III.D, Eqs. (8)-(9), Table IV]
    "Therefore, the mean lifetime is defined as: τ=τ_attempt<N>=τ_attempt/P_LZ = T/2P_LZ. (9) ... Table IV. Fitted kinetic parameters for the multi-passage survival cascade model."

    P_LZ and T in Table IV are fitted to the S2/S1/S0 population traces via Eqs. (2)-(7). Eq. (9) then defines the reported lifetimes as T/(2P_LZ). The 'channel-specific lifetimes' (13.6, 25.0, 222.2 fs) are therefore algebraic restatements of fitted parameters, not independent outputs from the first-principles dynamics. The paper's later statement that the model 'demonstrates high accuracy in reproducing the ultrafast population dynamics' is uninformative because the model parameters were optimized on those same curves.

  2. self definitional [Section III.D, Eq. (3), Table IV, and the concluding paragraph of Section III.D]
    "Importantly, a high-quality fit to the population dynamics is achievable only when the branching in S1 is explicitly included in the kinetic model. This requirement strongly supports the existence of the newly discovered photochemical pathway mediated by a novel σπ∗/S0 conical intersection – a mechanistic insight that would be nearly impossible to deduce from the population evolution alone without the present model."

    Eq. (3) already contains the second channel by construction: P_S1(t) = α·[channel a terms] + (1−α)·[channel b terms]. Fitting returns α=0.84, i.e. (1−α)=0.16, and this fitted branch is then offered as evidence that the σπ*/S0 pathway exists. No one-channel (α=1) comparison, information criterion, residual analysis, or cross-validation is reported. The 'requirement' that branching be included is asserted, not demonstrated; the fit's success is entailed by the model's own assumption of the pathway.

1 more flagged steps
  1. self citation load bearing [Introduction, Section III.C, Conclusions; ref. 38]
    "Building on our initial discovery of a UV-driven direct H2 elimination pathway in CH2NH2+ via a novel σπ∗/S0 conical intersection38 ... providing independent validation for the newly discovered photochemical pathway mediated by a novel σπ∗/S0 conical intersection38."

    The existence of the σπ*/S0 pathway is imported from the same group's preprint (ref. 38), which is also the stated source of the XMCQDPT2 reference MECI structures and of the trajectory-classification protocol. The paper calls the kinetic analysis an 'independent validation,' but that analysis assumes the pathway in Eq. (3) and then fits its branch weight. The load-bearing evidence for the central mechanistic claim therefore reduces to a self-citation plus a fit to a model that contains the pathway by construction, rather than to an external, independent test.

full rationale

The MLIP construction and the raw trajectory data are not circular: the SS-TL potentials are trained on XMCQDPT2 energies and forces, validated on out-of-trajectory splits, and tested against MECI geometries; the computed hopping times (5.0 fs and 28.4 fs) and the 2% H2-loss yield are internally consistent and comparable to previous CASSCF values. The circularity is confined to the kinetic-interpretation layer and to the claim that this layer independently validates the σπ*/S0 pathway. Eq. (9) makes each reported lifetime a closed-form function of P_LZ and T, which are themselves fitted to the population curves; presenting these lifetimes as 'extracted directly from first-principles simulations' and using the fit quality to 'validate the model physical picture' is a fitted-input-called-prediction pattern. Separately, the conclusion that the σπ*/S0 channel is supported rests on a branching term that is assumed in Eq. (3) and then fit; without any quantitative single-channel versus two-channel model comparison, the model's success is a consequence of its own construction. The self-citation to ref. 38 is load-bearing for the pathway's existence and for the 'independent validation' claim, although the paper does provide independent MECI-based evidence that the CI exists. I do not count the 'power-law' label as circularity, though it is internally inconsistent with Eqs. (2)-(7), since (1-P)^(c t) is exponential in t. Overall score 6: partial circularity, with the MLIP/dynamics core independent but the central mechanistic validation reducing to fitted parameters and model assumptions.

Assumptions & free parameters 10 free parameters · 7 assumptions · 1 invented entities

The central kinetic claim rests on a large set of fitted quantities: P_LZ, T, alpha, induction times, and dispersion parameters are all adjusted to match the simulated populations. Equation 9 then turns those fitted values into 'lifetimes,' so the model's quantitative reproduction of the target data is expected by construction. The electronic-structure and MLIP parts rest on the validity of XMCQDPT2 reference data, the LZBL hopping approximation, and the representativeness of the 1,450-geometry training set.

free parameters (10)
  • P_LZ^(1) (S2->S1 single-passage probability) = 0.92
    Fitted to population curves; Table IV.
  • P_LZ^(a) (pi-pi*/S0 path) = 0.60
    Fitted to population curves; Table IV.
  • P_LZ^(b) (sigma-pi*/S0 path) = 0.09
    Fitted to population curves; Table IV.
  • T1 (S2/S1 oscillation period) = 25 fs
    Fitted to population curves; Table IV.
  • Ta (pi-pi*/S0 oscillation period) = 30 fs
    Fitted to population curves; Table IV.
  • Tb (sigma-pi*/S0 oscillation period) = 40 fs
    Fitted to population curves; Table IV.
  • alpha (branching fraction to path a) = 0.84 (S2->S1 alpha = 1.00)
    Controls relative weight of the two S1-to-S0 channels; Table IV.
  • t_ind^(1), t_ind^(a), t_ind^(b) (induction times) = not reported
    Appear in Eqs. 6-7 but are absent from Table IV; presumably fitted or fixed.
  • sigma_T1, sigma_T_zeta (period dispersions) = 0.13, 0.20
    Chosen by hand to damp oscillations in Eq. 10.
  • Product-classification cutoff values = 3.55 A, 1.95 A, 1.64 A, 2.51 A, 1.95 A
    Established based on temporal variation in the same trajectories being classified; Methods II.G.
assumptions (7)
  • domain assumption XMCQDPT2/SA(3)-CASSCF(12,12)/aug-cc-pVDZ is treated as the reference truth for all three states.
    Methods II.A uses these energies as labels; if the reference method is inaccurate near seams, the MLIPs inherit its errors.
  • domain assumption LZBL surface hopping without explicit nonadiabatic couplings accurately captures the true nonadiabatic dynamics.
    Methods II.E; this is the mechanism that generates all population dynamics.
  • domain assumption The 1,450 sampled geometries cover the dynamically relevant configuration space for S2-initiated trajectories.
    Methods II.A; if trajectories explore unsampled regions, MLIP extrapolation errors are uncontrolled.
  • ad hoc to paper Each conical intersection passage is an independent Bernoulli trial with constant P_LZ and fixed period T, so populations follow Eqs. 2-7.
    Sec. III.D; this memoryless form contradicts the paper's description that the system 'remembers how many times it has attempted.'
  • ad hoc to paper Exactly two S1-to-S0 decay channels (pi-pi*/S0 and sigma-pi*/S0) with a fixed branching fraction alpha are sufficient to represent the decay.
    Eq. 3 and Table IV; no one-channel or alternative-model comparison is provided.
  • domain assumption The ensemble standard deviation of predicted energies is a valid uncertainty metric for trajectory weighting.
    Methods II.F; the corrected populations and hopping times depend on this assumption.
  • domain assumption Product-channel classification thresholds are objective and not biased by the outcomes they assign.
    Methods II.G; thresholds were set from the same trajectory set used for the branching-ratio results.
invented entities (1)
  • sigma-pi*/S0 conical intersection channel (path b) as a distinct S1 decay route
    purpose: Allows the kinetic model to fit a slow ~200 fs S1 decay component and supports the previously proposed H2-loss pathway.
    The only new evidence is a fit of a model that assumes this channel; the MECI itself was located in the authors' prior work (ref 38), and no experimental observable is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of XMCQDPT2-Fidelity Transfer-Learning Potentials and a Wavepacket Oscillation Model with Power-Law Decay for Ultrafast Photodynamics." pith.science (2026). https://pith.science/paper/X3PX4UNL

@misc{pith2026251207537,
  author       = {Pith},
  title        = {Pith review of: XMCQDPT2-Fidelity Transfer-Learning Potentials and a Wavepacket Oscillation Model with Power-Law Decay for Ultrafast Photodynamics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/X3PX4UNL}},
  note         = {Machine review of arXiv:2512.07537}
}
abstract

A central pursuit in theoretical chemistry is the accurate simulation of photochemical reactions, which are governed by nonadiabatic transitions through conical intersections. Machine learning has emerged as a transformative tool for constructing the necessary potential energy surfaces, but applying it to excited states faces a fundamental barrier: the cost of generating high-level quantum chemistry data. We overcome this challenge by developing machine-learning interatomic potentials (MLIPs) that achieve multi-state multi-reference perturbation theory accuracy through various techniques, such as transfer, multi-state, and $\Delta$-learning. Applied to the methaniminium cation, our highest-fidelity transfer-learning model uncovers its complete photodissociation landscape following S$_2$ photoexcitation. The comprehensive XMCQDPT2/SA(3)-CASSCF(12,12) electronic structure description captures all competing decay channels, including S$_1$ branching into photoisomerization and direct H$_2$-loss pathways. Our results show that the population dynamics generally depends on the MLIP model, correlating with its performance. At the same time, the introduction of MLIP-uncertainty corrections based on the predictions of an ensemble of models brings different approaches into agreement, validating this metric as essential for reliable dynamics. To interpret the population dynamics, we introduce a wavepacket oscillation model - a mechanistically transparent, power-law kinetics framework that extracts state-specific lifetimes directly from first-principles simulations. The model quantitatively reproduces the ultrafast decay, creating a direct link between quantum transition probabilities and classical rate constants. The kinetic fits yield channel-specific lifetimes, supporting the recently discovered photochemical pathway mediated by a novel $\sigma\pi^*/S_0$ conical intersection.

Figures

Figures reproduced from arXiv: 2512.07537 by the authors.

Figure 1
Figure 1. FIG. 1. Minimum-energy conical intersections (MECIs) located using an ensemble of the two best [PITH_FULL_IMAGE:figures/full_fig_p013_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2. Time evolution of electronic state populations for dynamics initiated in the S [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3. Branching ratios of the photodissociation channels after 100 fs, calculated using an ensem [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: FIG. 4. Time evolution of electronic state populations following photoexcitation to S [PITH_FULL_IMAGE:figures/full_fig_p020_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

5 extracted references · 1 canonical work pages

  1. [1]

    M¨ uller, S

    1C. M¨ uller, S. Srˇ seˇ n, B. Bachmair, R. Crespo-Otero, J. Li, S. Mausenberger, M. Pinheiro, G. Worth, S. A. Lopez, and J. Westermayr, Chem. Sci.16, 17542 (2025). 2J. Behler and M. Parrinello, Phys. Rev. Lett.98, 146401 (2007). 3R. Drautz, Phys. Rev. B99, 014104 (2019). 4C. Devereux, J. S. Smith, K. K. Huddleston, K. Barros, R. Zubatyuk, O. Isayev, and ...

    2025

  2. [2018]

    pp. 876–885. 45A. K. Belyaev and O. V. Lebedev, Phys. Rev. A84, 014701 (2011). 46Z. Varga, K. A. Parker, and D. G. Truhlar, Phys. Chem. Chem. Phys.20, 26643 (2018). 47Y. Li, J. Liu, J. Li, Y. Zhai, J. Yang, Z. Qu, and H. Li, J. Chem. Phys.155, 214102 (2021). 48J. O. Richardson, J. Chem. Phys.158, 011102 (2023). 49L. Zhang, S. V. Pios, M. Martyka, F. Ge, Y...

    2011

  3. [2021]

    9377–9388

    pp. 9377–9388. 6I. Batatia, D. P. Kovacs, G. Simm, C. Ortner, and G. Cs´ anyi, Adv. Neural Inf. Process. Syst.35, 11423 (2022). 7S. Batzner, A. Musaelian, L. Sun, M. Geiger, J. P. Mailoa, M. Kornbluth, N. Molinari, T. E. Smidt, and B. Kozinsky, Nat. Commun.13, 2453 (2022). 8J. Westermayr, M. Gastegger, D. V¨ or¨ os, L. Panzenboeck, F. Joerg, L. Gonz´ alez...

    arXiv 2022

  4. [2024]

    Firefly version 8,

    first edition ed., pp. 141–157. 19H. V. Kiefer, G. Elisabeth, L. Jeppe, P. A. Kusochek, A. V. Bochenkova, and L. H. Andersen, Nat. Commun.10, 1210 (2019). 20E. Gruber, A. M. Kabylda, M. B. Nielsen, A. P. Rasmussen, R. Teiwes, P. A. Kusochek, A. V. Bochenkova, and L. H. Andersen, J. Am. Chem. Soc.144, 69 (2021). 21A. Svendsen, H. V. Kiefer, H. B. Pedersen,...

    work page doi:10.48550/arxiv.2512.00913 2019

  5. [2025]

    Li, B.-B

    40S. Li, B.-B. Xie, B.-W. Yin, L. Liu, L. Shen, and W.-H. Fang, J. Phys. Chem. A128, 5516 (2024). 41Y. Huang, Y.-F. Hou, and P. O. Dral, Mach. Learn.: Sci. Technol.6, 035004 (2025). 42J. Westermayr, F. A. Faber, A. S. Christensen, O. A. von Lilienfeld, and P. Marquetand, Mach. Learn.: Sci. Technol.1, 025009 (2020). 43S. Mausenberger, C. M¨ uller, A. Tkatc...

    2024

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.