Pith. sign in

REVIEW 4 major objections 6 minor 51 references

A reinforcement-learning agent finds transition-state geometries by walking the energy surface from reactant to product, without labelled paths or expert guesses.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

Soft actor-critic RL locates first-order saddle points on realistic molecular PESs from reactant and product alone, without labelled paths or hand-crafted reaction coordinates.

T0 review reviewed 2026-07-14 challenge →

load-bearing objection Solid SAC agent for TS search on small GFN2-xTB systems; the “pathway-free” framing is a bit oversold because the product vector steers both actions and reward, but the engineering and benchmarks are real. the 4 major comments →

arxiv 2607.10854 v1 pith:XF2URLZC submitted 2026-07-12 physics.chem-ph

Autonomous Transition State Search with Soft Actor-Critic Reinforcement Learning

Utham Suresh , Konstantinos D. Vogiatzis This is my paper
classification physics.chem-ph
keywords transition state searchreinforcement learningsoft actor-criticpotential energy surfacesaddle pointinternal coordinatesautomated reaction discoverysemiempirical methods
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

Finding the transition state of a chemical reaction usually needs a good starting guess and expert intuition, or many expensive intermediate calculations along a guessed path. This paper shows that a soft actor-critic reinforcement-learning agent can locate those first-order saddle points by interacting directly with a realistic molecular energy surface. The agent starts from a known reactant, is told the product geometry, and at each step proposes changes to the molecule’s internal coordinates using only local energy, gradient, and curvature information. A multi-part reward pushes it toward regions with a single imaginary vibrational mode and low forces, while a progressive curriculum first explores broadly and then refines tightly. On eight standard benchmark reactions the resulting geometries match reference saddle points closely, often better than conventional path-interpolation methods that can fail or diverge. The result matters because it removes the need for labelled trajectories or hand-crafted reaction coordinates and therefore opens a route to more automated mapping of reaction networks.

Core claim

Without labelled trajectories or prescribed reaction pathways, a soft actor-critic agent that navigates a molecular potential energy surface from reactant toward product using only local gradients, Hessian eigenvalues, and a product-directed target vector can locate first-order saddle-point geometries that agree with established optimizers to chemical accuracy on standard benchmark reactions.

What carries the argument

The soft actor-critic transition-state search (SACTS) agent: a continuous-action policy that proposes internal-coordinate displacements as a learned combination of the softest Hessian mode, a gradient-steering direction, and a global step-size controller, trained by a multi-component reward that rewards progress toward the product, single-imaginary-mode regions, and low RMS gradient, with a two-phase curriculum that zooms from broad exploration into tight refinement once a candidate saddle zone is found.

Load-bearing premise

That a single product-directed target vector plus a hand-designed reward is enough to keep the agent on the chemically intended path when the molecule rearranges a lot, rather than letting the target overwhelm the curvature signals and send the agent to an unphysical or off-pathway saddle.

What would settle it

Run SACTS on a larger set of reactions that involve substantial geometric reorganization (for example multi-step rearrangements or ring openings without masking) and check whether the recovered saddles still match reference first-order saddle points to chemical accuracy, or whether they systematically land on competing pathways or unphysical geometries.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Transition-state searches can be started from reactant and product alone, without expert initial guesses or precomputed path interpolations.
  • The same agent architecture can be reused across chemically different systems once the product geometry is supplied, reducing system-specific tuning.
  • Candidate structures that fall short of strict convergence remain usable as high-quality starting points for conventional optimizers.
  • Progressive rendering (coarse exploration followed by local refinement) can cut the number of expensive electronic-structure evaluations needed to locate a saddle.
  • Reinforcement learning becomes a practical component of automated reaction-network exploration pipelines.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Transfer learning of a policy pretrained on small systems could amortize the cost of training a fresh agent for every new reaction.
  • Adding explicit connectivity or bonding descriptors to the state would help the agent distinguish early versus late transition states and competing pathways.
  • The same curriculum-plus-reward design could be tested on higher-level electronic-structure methods once the agent’s step budget is shown to remain modest.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The manuscript introduces SACTS, a soft actor-critic reinforcement-learning agent that locates first-order saddle points on molecular potential energy surfaces. Starting from a reactant geometry and knowing the product, the agent acts in continuous internal-coordinate space using a three-component action (soft-mode scaling, gradient-steering scaling, and a global step-size multiplier), a multi-component reward that encourages progress toward the product and a single-imaginary-mode region, and a two-phase progressive-rendering curriculum that tightens step sizes after a candidate TS zone is found. Electronic structure is evaluated on-the-fly with GFN2-xTB. On eight small benchmark reactions the method recovers geometries with RMSDs typically ≤0.04 Å relative to Berny references and outperforms linear and IDPP NEB on several cases where NEB fails or converges to the product. Code is deposited on GitHub.

Significance. If the central results hold under clearer framing, the work is a useful contribution to automated reaction-path discovery: it demonstrates that an off-policy maximum-entropy actor-critic agent can climb realistic molecular PESs using only local energy, gradient, and Hessian information, without labelled trajectories or a precomputed path of images. Strengths that should be credited include (i) direct interaction with a quantum-chemical environment rather than a pre-discretized grid or a pretrained MLIP, (ii) a continuous internal-coordinate action basis with progressive coarse-to-fine rendering, (iii) systematic comparison to Berny and NEB on the same surface, and (iv) public code. These elements establish a reproducible baseline for RL-based TS search even if transfer across systems and higher-level PESs remain future work.

major comments (4)
  1. Abstract and Action Representation: The claim that SACTS operates “without … prescribed reaction pathways” is overstated. The soft-mode eigenvector is sign-aligned by a positive projection onto the full reactant-to-product internal-coordinate displacement d_target, the same vector supplies the dense progress reward r_progress = ||d_target||_{t−1} − ||d_target||_t, and for ring opening the authors further mask spectator Z-matrix entries so that the prior remains chemically relevant. This is endpoint-conditioned geometric guidance of the same family used by double-ended methods. The manuscript should rephrase the abstract, introduction, and conclusions to state accurately that both endpoints are known and that a product-directed prior is built into the action basis and reward, while still distinguishing the approach from image-based path interpolation and from supervised trajectory learnin
  2. Action Representation and Results (HF + ethene; ring opening): Because d_target is load-bearing, the paper needs an ablation that quantifies how much performance depends on it. At minimum, report success rates, RMSDs, and episode counts with (i) the full target vector, (ii) the masked variant, and (iii) a pure soft-mode / gradient-steering policy with the progress term removed or replaced by a curvature-only objective. The HF-addition case (RMSD 0.19 Å, ΔE 7.63 kcal mol⁻¹) and the authors’ own remark that a large target vector can overwhelm curvature information make this ablation essential to support the claim of curvature-informed autonomous discovery rather than guided interpolation.
  3. Results and Conclusions: A new agent is trained from scratch for each reaction, yet the abstract and closing paragraphs present SACTS as a “general strategy” for “scalable, automated reaction discovery” and “transferable” search. The demonstrated scope is eight systems of ≤10 atoms on GFN2-xTB with per-system training. The claims should be scaled to what is shown, and any transfer or multi-task experiment (even a simple fine-tuning from one isomerisation to another) should be added or explicitly deferred as future work rather than implied by the present results.
  4. Results / Figure 4 and Computational Details: The NEB baselines use linear and IDPP interpolation without climbing-image refinement. The manuscript itself cites literature in which CI-NEB with internal-coordinate interpolation succeeds on comparable systems. Either add CI-NEB (or an equivalent strong double-ended baseline) under the same GFN2-xTB forces, or qualify the comparison so that SACTS is not presented as systematically superior to the best standard double-ended practice on these surfaces.
minor comments (6)
  1. Introduction: Several recent RL-for-PES and generative-TS papers are cited; a short explicit contrast table (grid Q-learning, actor-critic NEB, transfer-learning agents, SACTS) would help readers place the contribution.
  2. State Representation, Eq. (2)–(4): The repulsive proximity signal and the tanh-filtered eigenvalues are described only briefly; a one-sentence statement of units and typical numerical ranges would aid reproducibility.
  3. Reward Function: The main text defers the full formula to the SI. A compact equation or table of the additive terms and their phase-dependent coefficients in the main text would make the curriculum easier to follow without leaving the paper.
  4. Figure 2–3: Colour scales and axis labels are hard to read in the manuscript rendering; ensure high-contrast labels and a colour-blind-safe palette in the final figures.
  5. Throughout: Occasional typographical issues (e.g., “user s expertise”, spacing in author names, mixed en-dashes) should be cleaned in copy-editing.
  6. Data Availability: The GitHub link is welcome; please also archive a frozen commit or DOI (Zenodo/Figshare) so that the exact version used for the reported numbers remains citable after future code changes.

Circularity Check

0 steps flagged

No load-bearing circularity; empirical RL search uses endpoints by design and is validated against independent Berny/NEB references.

full rationale

The paper presents an empirical soft-actor-critic agent that interacts with a GFN2-xTB PES, using reactant and product geometries to construct a target displacement vector that signs the soft mode, supplies a dense progress reward, and appears in the state. This is transparent double-ended guidance, not a derivation that reduces the reported TS geometries to the inputs by construction. Success is measured by external criteria (RMSD and energy differences versus Berny-optimized saddles, plus NEB comparisons) and by the standard first-order saddle conditions (single negative Hessian eigenvalue, gradient norm < 0.005 Ha/Bohr). No parameters are fitted to the target TS data and then re-presented as predictions; no uniqueness theorem or ansatz is imported via self-citation to force the result; the reward is hand-designed to encourage the known topological signature of a first-order saddle, which is ordinary RL goal specification rather than circular self-definition. Minor self-citations to the authors’ prior ML reviews exist but are not load-bearing for the central claim. The method is therefore self-contained against external benchmarks; the only mild tension is rhetorical (abstract language of “no prescribed reaction pathways” versus the explicit product-directed vector), which does not constitute a circular reduction of equations or results.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 1 invented entities

The central claim rests on a large set of hand-chosen algorithmic hyperparameters (reward coefficients, phase step-size ceilings, convergence thresholds, SAC temperature target, network sizes) and on the modelling choice that GFN2-xTB plus a product-directed target vector adequately represents the chemistry. No new physical entities are postulated; the “invented” objects are the engineered RL components themselves.

free parameters (4)
  • phase-dependent step-size ceilings (bonds/angles/dihedrals)
    Explicitly set by hand (0.05/0.10/0.30 → 0.005/0.010/0.020) and reduced by an order of magnitude between phases; directly control exploration vs refinement.
  • reward coefficients (progress, perpendicular-gradient penalty, N_neg penalties, sparse TS-zone bonus, terminal rewards)
    Multiple additive terms whose relative weights are chosen by the authors and grow across phases; not derived from first principles.
  • SAC temperature target entropy and network architecture sizes
    Standard SAC free parameters; values relegated to ESI and affect exploration–exploitation balance.
  • gradient-norm and eigenvalue-count thresholds for phase transition and episode termination
    0.01 and 0.005 Hartree/Bohr, N_neg = 1 consistency checks; chosen to define “candidate zone” and “absolute convergence.”
axioms (4)
  • domain assumption A first-order saddle point is uniquely identified by exactly one negative Hessian eigenvalue and a near-zero gradient norm.
    Standard quantum-chemistry definition used throughout the reward and termination criteria.
  • domain assumption GFN2-xTB supplies a sufficiently realistic PES topology for the eight benchmark reactions.
    All forces, energies and Hessians are evaluated at this semiempirical level; higher-level QM is not used.
  • ad hoc to paper The softest Hessian eigenvector (sign-aligned with the reactant-to-product vector) plus the perpendicular gradient component form a useful continuous action basis.
    Core design choice of the action space; not forced by theory and noted by the authors to degrade for large reorganizations.
  • ad hoc to paper Stacking five consecutive observation vectors supplies adequate short-term memory for the policy.
    Adopted from Ahuja et al.; length of history is a free design decision.
invented entities (1)
  • SACTS agent (soft-actor-critic policy + critic trained with the specific multi-component reward and progressive-rendering curriculum) no independent evidence
    purpose: To map reactant geometries to first-order saddle points without labelled trajectories or fixed reaction coordinates.
    The composite architecture, reward shaping and two-phase curriculum are original to this paper; independent evidence is limited to the eight reported benchmarks.

reviewed 2026-07-14 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Autonomous Transition State Search with Soft Actor-Critic Reinforcement Learning." pith.science (2026). https://pith.science/paper/XF2URLZC

@misc{pith2026260710854,
  author       = {Pith},
  title        = {Pith review of: Autonomous Transition State Search with Soft Actor-Critic Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XF2URLZC}},
  note         = {Machine review of arXiv:2607.10854}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Transition state (TS) search is a crucial step in understanding chemical reactivity and mechanisms, yet conventional algorithms remain computationally intensive and heavily reliant on initial guesses, user s expertise, and chemical intuition. While recent machine learning approaches have shown promise, they demand either large training datasets or geometric interpolation between known endpoints, limiting their generality. In this work, we introduce a TS search model based on the soft actor-critic model, an advanced reinforcement learning algorithm in which an agent learns to navigate potential energy surfaces directly from local energetic and curvature information starting from a given reactant and its corresponding product. By formulating the search as a sequential decision-making process in internal coordinates, the agent adaptively proposes chemically meaningful structural updates through a reward function designed to promote movement towards saddle point regions. Without labelled trajectories or prescribed reaction pathways, the method successfully identifies TS geometries for standard benchmark reactions, operating directly on realistic molecular potential energy surfaces. These results highlight the potential of RL as a general strategy for reducing dependence on initial guesses and enabling scalable, automated reaction discovery across diverse chemical systems.

Figures

Figures reproduced from arXiv: 2607.10854 by Konstantinos D. Vogiatzis, Utham Suresh.

Figure 1
Figure 1. Figure 1: Schematic overview of the soft actor-critic transition state Search (SACTS) model. The actor proposes geometric displacements of the molecular structure (actions) based on the stacked-history state vector sₜ, while the critic evaluates the expected return to guide the actor toward first-order saddle points on the potential energy surface. State Representation The effectiveness of any reinforcement-based ex… view at source ↗
Figure 3
Figure 3. Figure 3: Reward per step for the HCN isomerization TS search with dark-to-light colour gradient maps progression from early to late episodes [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

51 extracted references · 2 linked inside Pith

  1. [1]

    G. N. Simm and M. Reiher, Journal of Chemical Theory and Computation, 2017, 13, 6108-6119

    2017

  2. [2]

    J. P . Unsleber, S. A. Grimmel and M. Reiher, Journal of Chemical Theory and Computation, 2022, 18, 5393-5409

    2022

  3. [3]

    J. Yin, W. Li, H. Chen, J. Qiu, H. Feng, X. Xu, Q. Jin and X. Wang, ACS Catalysis, 2025, 15, 15754-15764

    2025

  4. [4]

    S. M. Paul, D. S. Mytelka, C. T. Dunwiddie, C. C. Persinger, B. H. Munos, S. R. Lindborg and A. L. Schacht, Nature Reviews Drug Discovery, 2010, 9, 203-214

    2010

  5. [5]

    K. D. Vogiatzis, M. V . Polynski, J. K. Kirkland, J. Townsend, A. Hashemi, C. Liu and E. A. Pidko, Chemical Reviews, 2019, 119, 2453-2523

    2019

  6. [6]

    T. Das, S. Ghule and K. Vanka, ACS Central Science, 2019, 5, 1532-1540

    2019

  7. [7]

    J. P . Unsleber and M. Reiher, Annual Review of Physical Chemistry, 2020, 71, 121-142

    2020

  8. [8]

    Va z a r t , D

    F. Va z a r t , D. C a l d e r i n i , C. Pu z z a r i n i , D. S ko u t e r i s a n d V. B a r o n e, J o u r n a l o f C h e m i c a l Theory and Computation, 2016, 12, 5385-5397

    2016

  9. [9]

    J. S. Yoo, Y. Liu, X. Rong and A. M. Kolpak, The Journal of Physical Chemistry Letters, 2018, 9, 1473-1479

    2018

  10. [10]

    Prozument, J

    K. Prozument, J. H. Baraban, P . B. Changala, G. B. Park, R. G. Shaver, J. S. Muenter, S. J. Klippenstein, V . Y . Chernyak and R. W. Field, Proceedings of the National Academy of Sciences, 2020, 117, 146-151

    2020

  11. [11]

    L i u , D

    Y. L i u , D. M . S a n c h e z , M . R . Wa r e, E . G. C h a m p e n o i s , J. Ya n g , J. P. F. N u n e s , A . A t t a r, M . Centurion, J. P . Cryan, R. Forbes, K. Hegazy, M. C. Ho]mann, F . Ji, M. F . Lin, D. Luo, S. K. Saha, X. Shen, X. J. Wang, T. J. Martínez and T. J. A. Wolf, Nature Communications, 2023, 14,

    2023

  12. [12]

    H. B. Schlegel, WIREs Computational Molecular Science, 2011, 1, 790-809

    2011

  13. [13]

    H. B. Schlegel, Journal of Computational Chemistry, 1982, 3, 214-218

    1982

  14. [14]

    Maeda, Y

    S. Maeda, Y . Harabuchi, M. Takagi, T. Taketsugu and K. Morokuma, The Chemical Record, 2016, 16, 2232-2248

    2016

  15. [15]

    Henkelman, B

    G. Henkelman, B. P . Uberuaga and H. Jónsson, The Journal of Chemical Physics, 2000, 113, 9901-9904

    2000

  16. [16]

    Peters, A

    B. Peters, A. Heyden, A. T. Bell and A. Chakraborty, The Journal of Chemical Physics, 2004, 120, 7877-7886

    2004

  17. [17]

    Z. D. Pozun, K. Hansen, D. Sheppard, M. Rupp, K.-R. Müller and G. Henkelman, The Journal of Chemical Physics, 2012,

    2012

  18. [18]

    A. A. Peterson, The Journal of Chemical Physics, 2016,

    2016

  19. [19]

    Denzel and J

    A. Denzel and J. Kastner , Journal of Chemical Theory and Computation, 2018, 14, 5777-5786

    2018

  20. [20]

    Meyer, K

    R. Meyer, K. S. Schmuck and A. W. Hauser, Journal of Chemical Theory and Computation, 2019, 15, 6513-6523

    2019

  21. [21]

    Pattanaik, J

    L. Pattanaik, J. B. Ingraham, C. A. Grambow and W. H. Green, Physical Chemistry Chemical Physics, 2020, 22, 23618-23626

    2020

  22. [22]

    Jackson, W

    R. Jackson, W. Zhang and J. Pearson, Chemical Science, 2021, 12, 10022-10040

    2021

  23. [23]

    G. M. Jones, B. A. Smith, J. K. Kirkland and K. D. Vogiatzis, Inorganic Chemistry Frontiers, 2023, 10, 1062-1075

    2023

  24. [24]

    Q. Zhao, Y. Han, D. Zhang, J. Wang, P . Zhong, T. Cui, B. Yin, Y. Cao, H. Jia and C. Duan, Advanced Science, 2025, 12, e06240. 20

    2025

  25. [25]

    K. D. Vogiatzis, C. Corminboeuf, A. Nova, K. Jorner, J. Kästner, M. Meuwly, P . Schwaller, V . Böttcher, M. Drosou, E. Fako, H. Hoppe, Z. Ivkovic, N. Iwanojko, D. A. Pantazis, S. P . Schmid, K. Szenes, A. Tetenoire and M. Reiher, Journal of the American Chemical Society, 2026, 148, 9143-9155

    2026

  26. [26]

    Käser, L

    S. Käser, L. I. Vazquez-Salazar, M. Meuwly and K. Töpfer, Digital Discovery, 2023, 2, 28-58

    2023

  27. [27]

    T. Cui, Y. Han, H. Jia, C. Duan and Q. Zhao, Scientific Data, 2025, 13,

    2025

  28. [28]

    Z. Wu, L. Zhou, P . Hou, Y . Liu, R. Wang, T. Guo and J.-C. Liu, JACS Au, 2025, 5, 6151-6161

    2025

  29. [29]

    M. Z. Makoś, N. Verma, E. C. Larson, M. Freindorf and E. Kraka, The Journal of Chemical Physics, 2021,

    2021

  30. [30]

    C. Duan, Y. Du, H. Jia and H. J. Kulik, Nature Computational Science, 2023, 3, 1045-1055

    2023

  31. [31]

    S. Kim, J. Woo and W. Y . Kim, Nature Communications, 2024, 15,

    2024

  32. [32]

    Duan, G.-H

    C. Duan, G.-H. Liu, Y . Du, T. Chen, Q. Zhao, H. Jia, C. P . Gomes, E. A. Theodorou and H. J. Kulik, Nature Machine Intelligence, 2025, 7, 615-626

    2025

  33. [33]

    Galustian, K

    L. Galustian, K. Mark, J. Karwounopoulos, M. P . P . Kovar and E. Heid, Digital Discovery, 2025, 4, 3492-3501

    2025

  34. [34]

    S. Gow, M. Niranjan, S. Kanza and J. G. Frey, Digital Discovery, 2022, 1, 551-567

    2022

  35. [35]

    M n i h , K

    V. M n i h , K . K a v u kc u o g l u , D. S i l v e r, A . G ra v e s, I . A n t o n o g l o u , D. W i e r s t ra a n d M . A . Riedmiller, ArXiv, 2013, abs/1312.5602

    Pith/arXiv arXiv 2013

  36. [36]

    A. W. Mills, J. J. Goings, D. Beck, C. Yang and X. Li, Journal of Chemical Information and Modeling, 2022, 62, 3169-3179

    2022

  37. [37]

    Pal, Data Science, 2024, 7, 73-92

    A. Pal, Data Science, 2024, 7, 73-92

    2024

  38. [38]

    W. Xu, Y. Zhao, J. Chen, Z. Wan, D. Yan, X. Zhang and R. Zhang, Journal of Computational Chemistry, 2024, 45, 487-497

    2024

  39. [39]

    Barrett and J

    R. Barrett and J. Westermayr, The Journal of Physical Chemistry Letters, 2024, 15, 349-356

    2024

  40. [40]

    K. T. Schütt, O. T. Unke and M. Gastegger, Proceedings of the 38th International Conference on Machine Learning, 2021, 139, 9377-9388

    2021

  41. [41]

    Kikkawa, T

    T. Kikkawa, T. Kawakami, S. Yamanaka and M. Okumura, Chemistr y Letters, 2025,

    2025

  42. [42]

    Terven, AI, 2025, 6,

    J. Terven, AI, 2025, 6,

    2025

  43. [43]

    R. S. Sutton, D. Precup and S. Singh, Artificial Intelligence, 1999, 112, 181-211

    1999

  44. [44]

    M. E. Ororbia and G. P . Warn, Journal of Computing and Information Science in Engineering, 2021,

    2021

  45. [45]

    Z. Hou, K. Zhang, Y . Wan, D. Li, C. Fu and H. Yu, ArXiv, 2020, 2002.02829

    Pith/arXiv arXiv 2020

  46. [46]

    Haarnoja, A

    T. Haarnoja, A . Zhou, P. Abbeel and S. Levine, Proceedings of the 35th International Conference on Machine Learning, 2018, 1861-1870

    2018

  47. [47]

    Bannwarth, E

    C. Bannwarth, E. Caldeweyher, S. Ehlert, A. Hansen, P . Pracht, J. Seibert, S. Spicher and S. Grimme, WIREs Computational Molecular Science, 2021, 11, e1493

    2021

  48. [48]

    Hjorth Larsen, J

    A. Hjorth Larsen, J. Jørgen Mortensen, J. Blomqvist, I. E. Castelli, R. Christensen, M. Dułak, J. Friis, M. N. Groves, B. Hammer, C. Hargus, E. D. Hermes, P . C. Jennings, P . Bjerre Jensen, J. Kermode, J. R. Kitchin, E. Leonhard Kolsbjerg, J. Kubal, K. Kaasbjerg, S. Lysgaard, J. Bergmann Maronsson, T. Maxson, T. Olsen, L. Pastewka, A. Peterson, C. Rostga...

    2017

  49. [49]

    Baker and F

    J. Baker and F. Chan, Journal of Computational Chemistry, 1996, 17, 888-904

    1996

  50. [50]

    A. B. Birkholz and H. B. Schlegel, Journal of Computational Chemistry, 2015, 36, 1157-1166

    2015

  51. [51]

    Smidstrup, A

    S. Smidstrup, A. Pedersen, K. Stokbro and H. Jónsson, The Journal of Chemical Physics, 2014,

    2014

This paper was first reviewed by grok-4.5 on July 14, 2026.