Pith. sign in

REVIEW 4 major objections 6 minor 41 references

Acceleration of crystal structure relaxation with Deep Reinforcement Learning

T0 review · 4 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

Pith's one-line read A deep reinforcement-learning agent trained on crystal graphs can relax Al-Fe structures in fewer optimizer steps than classical solvers for larger unit cells, and its policy transfers across compositions and supercells.

desk verdict Careful proof-of-concept for RL relaxation, but the 'acceleration' claim rests on memorized benchmarks and missing wall-clock costs. read the letter →

arxiv 2502.08405 v1 pith:VVLCHNOC submitted 2025-02-12 cond-mat.mtrl-sci

classification cond-mat.mtrl-sci
keywords structurerelaxationreinforcementlearningE(3)-equivariantneuralnetworkstensorfieldTD3crystalgraphsAl-Fealloysembedded-atommethod
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

Structure relaxation—moving atoms toward a local energy minimum—is the routine bottleneck of atomistic simulation, and this paper asks whether a reinforcement-learning agent can learn to do it with fewer optimizer steps than classical methods. The authors frame relaxation as a Markov decision process on crystal graphs and train a Twin Delayed DDPG (TD3) agent whose actor and critic are E(3)-equivariant Tensor Field Networks, using forces from embedded-atom-method (EAM) potentials for Al-Fe systems. Their central quantitative finding is that for unit cells with 10 and 20 atoms the trained agent reaches force thresholds of 0.2 and 0.25 eV/Å in fewer steps than BFGS or conjugate gradient, while for smaller cells it is comparable. They also report that the learned policy generalizes: a model trained only on monatomic Al and Fe relaxes Al-Fe compounds, and models trained on small cells partially relax larger supercells and 4–5 atom compounds. The paper argues this makes DRL suitable as a baseline optimizer for repeatedly relaxing similar structures, such as catalytic surfaces, rather than as a replacement for classical optimizers at high convergence.

What carries the argument

The mechanical core is the equation of structure relaxation with a Markov decision process whose states are crystal graphs and whose actions are per-atom displacement vectors, together with an E(3)-equivariant Tensor Field Network used for both actor and critic. Equivariance here means that rotating or translating the input crystal graph rotates the predicted displacement field correspondingly, which by Curie's principle confines the policy to actions of equal or higher symmetry than the structure and removes most of the action space from consideration. Training uses TD3 with exploration applied by adding noise to the state's force vectors rather than to actions, plus an additional greedy-exploration rule that fires when forces are small but shifts are not; reward shaping combines force, log-force, and step rewards, and the force threshold $\epsilon_1$ is lowered sequentially during training so that the agent encounters terminal states early. This combination—symmetry-constrained actions, state noise, greedy resets, and staged reward shaping—is what carries the step-count advantage.

What would settle it

Run the trained TD3 agent and BFGS on the same N=20 Al-Fe structure to the same force threshold on the same hardware and compare total wall-clock time; if the agent takes longer in seconds despite fewer steps, the acceleration claim fails. A complementary check would swap the EAM environment for DFT forces on a small cell and see whether the step-count advantage persists when the environment is noisier and each evaluation is expensive.

Watch

Extended reading notes

Core claim

The discovery, stated on the paper's own terms, is that structure relaxation can be learned as a sequential decision problem: at each step the agent reads a crystal graph whose node features include the current forces, predicts a displacement vector for every atom, and receives a reward tied to the largest residual force or to the number of steps taken. The authors find the choice of network architecture decisive: a plain graph convolutional network fails, while an E(3)-equivariant tensor field network succeeds because it restricts predicted shifts to the subspace compatible with the structure's symmetry, shrinking the action space and avoiding rotation augmentation. With this architecture, using TD3 and noise-to-state exploration, the trained agent relaxes the CsCl-type AlFe structure in about 11 steps near its best setting and, for $N=10$ and $N=20$ unit cells, outperforms BFGS and conjugate gradient in step count to the moderate thresholds $0.2$ and $0.25$ eV/Å. The same method reaches only $0.1$–$0.01$ eV/Å reliably, and below that the model loses sensitivity near the minimum; the paper therefore positions the method as a fast first-stage optimizer. Transfer experiments show the model generalizes learned interaction patterns from Al-Al and Fe-Fe to Al-Fe and from smaller to larger cells, with the caveat that only a fraction of 4–5 atom structures relax within practical thresholds.

Load-bearing premise

Fewer relaxation steps is treated as equivalent to faster relaxation: the paper reports only step counts and never wall-clock time or per-step cost, so if one policy-network evaluation costs more than one classical energy-and-force call, the claimed acceleration would not survive.

Editorial extensions

If this is right

  • For repeated relaxation of structurally related configurations, a trained agent can replace most of the classical optimizer's early iterations, cutting the number of force evaluations needed to reach $\epsilon_1 \approx 0.1$–$0.01$ eV/Å.
  • The method is best deployed as a hybrid: let the RL agent relax to moderate forces, then hand the nearly quadratic residual to BFGS or conjugate gradient for final convergence.
  • Equivariance is a practical design rule for RL in materials: models built with E(3)-equivariant networks learn faster and need no rotation augmentation, while symmetry-breaking exploration noise is counterproductive.
  • Transfer along composition and system size is real but bounded: current models trained on 2–3 atom cells relax only about 10–40% of unseen 4–5 atom structures to practical thresholds, so curriculum training on progressively larger structures is the indicated path.

Reading between the lines

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

  • Editorial inference: whether fewer steps means less wall-clock time is untested here; a policy-network forward pass likely costs more than one EAM force call, so the acceleration case is stronger when the environment uses expensive ab initio forces such as DFT rather than a cheap classical potential.
  • Editorial inference: the reported sensitivity failure near minima suggests a concrete two-phase scheduler—RL to $\epsilon_1 \sim 0.1$ eV/Å, then a classical local optimizer—could deliver both the step-count gain and tight convergence; benchmarking this hybrid's total elapsed time on a fixed GPU/CPU would settle the acceleration question.
  • Editorial inference: the transfer results imply a curriculum/active-learning loop—starting with 2–3 atom cells and adding unseen environments only when success fraction stalls—could raise the success rate on 4–5 atom compounds, a direction the authors mention but do not test.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper introduces a deep reinforcement learning (DRL) approach to crystal structure relaxation, representing structures as crystal graphs and using the TD3 algorithm with TFN or CGCNN policies. Experiments on Al-Fe systems with EAM potentials compare architectures, reward functions, exploration strategies, and hyperparameters, and benchmark the best models against BFGS and conjugate gradient in terms of the number of relaxation steps. The paper also tests generalization to supercells and chemically more diverse structures. The central claim is that DRL models can enhance the efficiency of structure relaxation compared to classical optimizers.

Significance. If the efficiency claim were established with wall-clock measurements and a fair baseline, this would be a useful proof-of-concept for applying equivariant reinforcement learning to structure relaxation, and the detailed ablation of reward functions, exploration methods, and architectures would be valuable to the community. The paper is transparent: code, data, and hyperparameters are provided, and important limitations (sensitivity near minima, poor success rates on unseen structures) are stated explicitly. However, the central 'acceleration' claim is not currently supported by the reported step-count-only comparison.

major comments (4)
  1. [Section 3.3, Fig. 5a] The claim that DRL 'accelerates' relaxation is supported only by a comparison of the number of relaxation steps; no wall-clock time, CPU/GPU time, or per-step cost is reported anywhere in the paper. Each RL rollout step involves a TFN policy forward pass plus an ASE/pymatgen environment update with an EAM force evaluation, whereas each BFGS/CG step involves one or more EAM evaluations plus a cheap numerical update. Since EAM potentials are inexpensive for N=10–20 atoms, the TFN inference could dominate the per-step cost, and a 2–3x step reduction would not necessarily translate into faster relaxation. Without a cost comparison, the title's 'acceleration' and the abstract's 'enhance the efficiency' are unsupported inferences.
  2. [Section 3.3] The main benchmark trains the agent on the same structure it later relaxes: 'For each N the models are trained on a single structure and tested to perform the relaxation of the same structure but randomly distorted at the beginning of the testing episode.' Classical optimizers see that structure for the first time. Even for the intended use case of repeated relaxation of the same structure (Section 1), a fair comparison would give the classical optimizer access to prior relaxations (e.g., warm-starting from the previous final structure or Hessian). As reported, Fig. 5a conflates the benefit of pre-training on the exact PES with the algorithmic efficiency of the learned policy. A held-out benchmark or a warm-started classical baseline is needed to support the comparison.
  3. [Section 3.3 / Fig. 5a caption] The force threshold ϵ1 varies with system size: N=2–8 are relaxed to 0.01 eV/Å, while N=10 and N=20 are relaxed only to 0.2 and 0.25 eV/Å, respectively. The apparent advantage for larger N is therefore confined to moderate force convergence, and the paper's own conclusion (Section 4) recommends switching to classical optimizers below 0.1–0.01 eV/Å. The efficiency claim should be scoped to this regime, or additional results at tighter thresholds should be provided.
  4. [Section 3.4.3, Fig. 7] The generalization experiment most relevant to practical use reports success fractions of only 0.1–0.4 at ϵ1≈0.1 eV/Å for unseen structures, which the authors themselves describe as 'not a sufficient outcome.' This undercuts the broader claim that the model 'can be applied to relax more complex structures without pre-training on them' and should be explicitly integrated into the conclusions and abstract, which currently state a more general potential for efficiency enhancement.
minor comments (6)
  1. [Section 2.3] Typo in the subsection heading: 'Tensor-Filed' should be 'Tensor Field'.
  2. [Appendix C] Typos: 'demostrate' should be 'demonstrate', and 'exhausting experience' should be 'exhaustive experience'.
  3. [Appendix D, Fig. 8 caption] Typo in the caption: 'relaxion' should be 'relaxation'.
  4. [Eq. (13)] The expression for the geometric sum is ambiguous as typeset; please write it as (1−γ^n)/(γ−1) to avoid confusion.
  5. [Section 3.4.3] The 'normalized last step' metric n·ϵ1 is introduced without justification; please explain why multiplying the number of steps by the force threshold is a meaningful normalization across different ϵ1 values.
  6. [Appendix K] The statement 'start step was 500' is unclear; please specify what 'start step' refers to (e.g., number of warm-up environment steps before learning begins).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the in-sample benchmark and missing wall-clock timing are evidence limitations, not circular steps.

full rationale

The paper contains no formal derivation chain whose outputs are equivalent to its inputs. Its central comparison (Section 3.3, Fig. 5a) is an empirical benchmark: agents are trained on a single structure and then evaluated on random distortions of that same structure. This is an in-sample evaluation, but the paper does not present it as a generalization prediction; it explicitly frames the intended use case as 'relaxation of identical structures from different unstable initial states' (Conclusions), for which training on the target structure is appropriate rather than circular. Cross-structure generalization is separately tested out-of-sample in Section 3.4 (monoatomic Al/Fe to AlFe; small N to supercells; N=2-3 to N=4-5), and these tests are not reducible to training data by construction. The only load-bearing citations to external results (e.g., Curie's principle from Smidt et al. [37]) are used to motivate architecture choice, not to define the measured performance. No parameter is fitted and then renamed as a prediction; no equation is defined in terms of the quantity it is said to predict. The title's 'acceleration' claim is weakened by the absence of wall-clock or per-step cost data, and the in-sample benchmark limits the external generality of the headline numbers, but these are correctness and evidence concerns, not circularity. There is also no self-citation chain that forces the conclusion. The empirical work is therefore self-contained as reported, with the stated caveats.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The paper's central results depend on a set of hand-chosen thresholds (force criterion, distortion level, reward weights, exploration noise, discount factor) and on domain assumptions about the EAM potential and the graph representation. The TFN model itself contains thousands of fitted neural network weights, but those are trained rather than free in the ledger sense. No new physical entities are introduced.

free parameters (6)
  • Force threshold epsilon1 = varies from 1e-6 to 0.25 eV/A depending on experiment
    Defines the done criterion and directly controls the number of steps; different thresholds are used for different N, making cross-system comparisons inconsistent.
  • Distortion parameter beta = 0.5 (0.05 in Fig. 2-3)
    Controls the starting point of each episode; the paper finds the model can extrapolate to higher beta, but all training uses beta=0.5.
  • Reward weights (w1, w2, w3) = 1, 0, 0.5 (hybrid)
    Hand-chosen composition of force, step, and log-force rewards in Eq. (5); the central results depend on this choice.
  • Discount factor gamma = 0.9, 0.99, or 0.9999
    Tuned per experiment; affects stability and step count, as shown in Appendix H.
  • Noise level lambda = [0.01, 0.001], 0.2, or [0.2, 0.02]
    Exploration noise added to state; the schedule is hand-tuned and significantly affects performance.
  • Greedy exploration parameters (Ngr, delta_rmax, fmax) = Not stated in the main text
    Introduced in Section 2.4 to avoid getting stuck, but actual values are not reported.
assumptions (5)
  • domain assumption The EAM potential accurately represents Al-Fe interatomic interactions for the purpose of comparing optimizers.
    All PES evaluations use EAM; the paper does not validate against DFT or other potentials, so the relaxation landscape is an EAM landscape.
  • domain assumption The crystal graph with a 5 A cutoff captures all relevant local interactions for relaxation.
    Graph construction in Appendix A.3 uses rmax=5 A; interactions beyond this distance are neglected.
  • standard math The relaxation task satisfies the Markov property when states are crystal graphs with force features.
    Section 2.1 assumes st+1 depends only on st and at, which is plausible for deterministic force fields but not discussed.
  • domain assumption The maximum force criterion is a sufficient convergence measure for relaxation quality.
    Used in Eq. (6) as the done condition; the paper never relates this to energy error or structural accuracy.
  • standard math TFN equivariance plus Curie's principle restricts actions to a symmetry-appropriate subspace.
    Assumed in Section 3.1 to explain why TFN outperforms CGCNN.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Acceleration of crystal structure relaxation with Deep Reinforcement Learning." pith.science (2026). https://pith.science/paper/VVLCHNOC

@misc{pith2026250208405,
  author       = {Pith},
  title        = {Pith review of: Acceleration of crystal structure relaxation with Deep Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VVLCHNOC}},
  note         = {Machine review of arXiv:2502.08405}
}
read the original abstract

We introduce a Deep Reinforcement Learning (DRL) model for the structure relaxation of crystal materials and compare different types of neural network architectures and reinforcement learning algorithms for this purpose. Experiments are conducted on Al-Fe structures, with potential energy surfaces generated using EAM potentials. We examine the influence of parameter settings on model performance and benchmark the best-performing models against classical optimization algorithms. Additionally, the model's capacity to generalize learned interaction patterns from smaller atomic systems to more complex systems is assessed. The results demonstrate the potential of DRL models to enhance the efficiency of structure relaxation compared to classical optimizers.

Figures

Figures reproduced from arXiv: 2502.08405 by the authors.

Figure 1
Figure 1. Workflow of the Reinforcement Learning algorithm used for structure relaxation. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Learning curves of TD3 Agents with CGCNN and TFN architectures, trained on CsCl-type structure of [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. a) - b) Learning curves of TD3 Agents with different reward functions, given by Eq. (2)-(4). The Agents [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: Comparison of classical algorithms and TD3 Agents trained with and without switching to the [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: a) Comparison of TD3 model and classical optimizers in relaxation of structures with different number of [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Number of steps required for relaxation of supercells by TD3 Agents trained on structures with different [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Learning curve of the TD3 model, trained on random Al and Fe configurations with [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: a) Maximum norm of the force maxn∈[1,N] | ⃗fn(s ∗ )| in the structure s ∗ , distorted from the local minimum s0 by shifting the first atom in the structure to the vector {∆x, ∆y, 0}. The force is in eV/A; b) Logarithm of the norm of the difference in the model’s predic…
Figure 9
Figure 9. Figure 9: Learning curves of TD3 Agents with different reward functions, given by Eq. (2)-(4) with [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]
Figure 10
Figure 10. Figure 10: Several relaxation trajectories, conducted by the Agent trained on hypothetical I4/mmm structure of Al, are [PITH_FULL_IMAGE:figures/full_fig_p016_10.png]
Figure 11
Figure 11. Figure 11: Comparison of TD3 Agents trained with ("Greedy (force)" and "Greedy (hybrid)") and without ("Ordinary") [PITH_FULL_IMAGE:figures/full_fig_p017_11.png]
Figure 12
Figure 12. Figure 12: Performance of the best in terms of relaxation steps of TD3 and SAC models (marked with circles in Fig. [PITH_FULL_IMAGE:figures/full_fig_p018_12.png]
Figure 13
Figure 13. Figure 13: Learning curves comparing the performance of TD3 Agents with different discount factors [PITH_FULL_IMAGE:figures/full_fig_p019_13.png]
Figure 14
Figure 14. Figure 14: Learning curves comparing the performance of TD3 Agents with different discount factors [PITH_FULL_IMAGE:figures/full_fig_p020_14.png]
Figure 15
Figure 15. Figure 15: Learning curves comparing the performance of SAC Agents with different discount factors [PITH_FULL_IMAGE:figures/full_fig_p021_15.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 16 canonical work pages

  1. [1]

    Ionic Minimization in VASP

    Georg Kresse, Martijn Marsman, Michael Poeltl, Merzuk Kaltak, Christian Hirsch, and Ferenc Karsai. Ionic Minimization in VASP. https://www.vasp.at/wiki/index.php/Category:Ionic_minimization

  2. [2]

    Convergence acceleration of iterative sequences

    Péter Pulay. Convergence acceleration of iterative sequences. The case of SCF iteration. Chemical Physics Letters, 73(2):393–398, 1980. ISSN: 0009-2614. DOI: 10.1016/0009-2614(80)80396-4

  3. [3]

    Press, B

    William H. Press, B. P. Flannery, S. A. Teukolsky, and W. T. Vetterling.Numerical Recipes: The Art of Scientific Computing. Cambridge University Press, Cambridge, MA, 1st edition, 1986

  4. [4]

    Tschopp, and Srikanth Patala

    Arash Dehghan Banadaki, Mark A. Tschopp, and Srikanth Patala. An efficient Monte Carlo algorithm for determining the minimum energy structures of metallic grain boundaries. Computational Materials Science , 155:466–475, 2018. DOI: 10.1016/j.commatsci.2018.09.017

  5. [5]

    Pillardy, Y

    J. Pillardy, Y . A. Arnautova, C. Czaplewski, K. D. Gibson, and H. A. Scheraga. Conformation-family Monte Carlo: A new method for crystal structure prediction. Proceedings of the National Academy of Sciences, 98(22):12351– 12356, 2001. DOI: 10.1073/pnas.231479298

  6. [6]

    Parrinello and A

    M. Parrinello and A. Rahman. Crystal structure and pair potentials: A molecular-dynamics study. Phys. Rev. Lett., 45(14):1196–1199, 1980. American Physical Society. DOI: 10.1103/PhysRevLett.45.1196

  7. [7]

    Podryabinkin, Evgeny V

    Evgeny V . Podryabinkin, Evgeny V . Tikhonov, Alexander V . Shapeev, and Artem R. Oganov. Accelerating crystal structure prediction by machine-learning interatomic potentials with active learning. Phys. Rev. B, 99(6):064114, February 2019. American Physical Society. DOI: 10.1103/PhysRevB.99.064114. 21 A PREPRINT - F EBRUARY 13, 2025 Table 5: Structures us...

  8. [8]

    A low resources space time approach to the GW approximation

    Dietrich Foerster and Saber Gueddida. A low resources space time approach to the GW approximation. Computa- tional Materials Science, 187:110078, 2021. DOI: 10.1016/j.commatsci.2020.110078

Show all 41 references
  1. [9]

    Strout and Gustavo E

    Douglas L. Strout and Gustavo E. Scuseria. A quantitative study of the scaling properties of the Hartree–Fock method. The Journal of Chemical Physics, 102(21):8448–8452, June 1995. AIP Publishing. DOI: 10.1063/1.468836

  2. [10]

    Maja-Olivia Lenz, Thomas A. R. Purcell, David Hicks, Stefano Curtarolo, Matthias Scheffler, and Christian Carbogno. Parametrically constrained geometry relaxations for high-throughput materials science. Computational Materials, 5:123, 2019. DOI: 10.1038/s41524-019-0254-4

  3. [11]

    Deep reinforcement learning for inverse inorganic materials design

    Elton Pan, Christopher Karpovich, and Elsa Olivetti. Deep reinforcement learning for inverse inorganic materials design. npj Computational Materials, 10:287, 2024. DOI: 10.1038/s41524-024-01474-5

  4. [12]

    Gu, and Liwei Lin

    Fanping Sui, Ruiqi Guo, Zhizhou Zhang, Grace X. Gu, and Liwei Lin. Deep reinforcement learning for digital materials design. ACS Materials Letters, 3(10):1433–1439, 2021. DOI: 10.1021/acsmaterialslett.1c00390

  5. [13]

    Learning in continuous action space for developing high dimensional potential energy models

    Sukriti Manna, Troy Loeffler, Rohit Batra, Suvo Banik, Henry Chan, Bilvin Varughese, Kiran Sasikumar, Michael Sternberg, Tom Peterka, Mathew Cherukara, Stephen Gray, Bobby Sumpter, and Subramanian Sankaranarayanan. Learning in continuous action space for developing high dimens...

  6. [14]

    Zare, and Patrick Riley

    Zhenpeng Zhou, Steven Kearnes, Li Li, Richard N. Zare, and Patrick Riley. Optimization of molecules via deep reinforcement learning. Scientific Reports, 9(1):10752, July 2019. Springer Science and Business Media LLC. DOI: 10.1038/s41598-019-47148-x

  7. [15]

    Coley, Yiming Mo, Regina Barzilay, and Klavs F

    Xiaoxue Wang, Yujie Qian, Hanyu Gao, Connor W. Coley, Yiming Mo, Regina Barzilay, and Klavs F. Jensen. Towards efficient discovery of green synthetic pathways with Monte Carlo tree search and reinforcement learning. Chemical Science, 11(40):10959–10972, 2020. Royal Society of ...

  8. [16]

    Vasudevan, Erick Orozco, and Sergei V

    Rama K. Vasudevan, Erick Orozco, and Sergei V . Kalinin. Discovering mechanisms for materials microstructure optimization via reinforcement learning of a generative model. Machine Learning: Science and Technology , 3(4):04LT03, 2022. IOP Publishing. DOI: 10.1088/2632-2153/aca004

  9. [17]

    Grossman

    Tian Xie and Jeffrey C. Grossman. Crystal graph convolutional neural networks for an accurate and interpretable prediction of material properties. Physical Review Letters, 120(14), 2018. American Physical Society. DOI: 10.1103/physrevlett.120.145301. 22 A PREPRINT - F EBRUARY ...

  10. [18]

    Tensor field networks: Rotation- and translation-equivariant neural networks for 3D point clouds

    Nathaniel Thomas, Tess Smidt, Steven Kearnes, Lusann Yang, Li Li, Kai Kohlhoff, and Patrick Riley. Tensor field networks: Rotation- and translation-equivariant neural networks for 3D point clouds. arXiv preprint, 2018. arXiv:1802.08219

  11. [19]

    Addressing function approximation error in actor-critic methods

    Scott Fujimoto, Herke van Hoof, and David Meger. Addressing function approximation error in actor-critic methods. arXiv preprint, 2018. arXiv:1802.09477

  12. [20]

    Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor

    Tuomas Haarnoja, Aurick Zhou, Pieter Abbeel, and Sergey Levine. Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor. arXiv preprint, 2018. arXiv:1801.01290

  13. [21]

    G. J. Ackland, D. J. Bacon, A. F. Calder, and T. Harry. Computer simulation of point defect properties in dilute Fe—Cu alloy using a many-body interatomic potential. Philosophical Magazine A, 75(3):713–732, 1997. DOI: 10.1080/01418619708207198

  14. [22]

    Zope and Y

    Rajendra R. Zope and Y . Mishin. Interatomic potentials for atomistic simulations of the Ti-Al system.Phys. Rev. B, 68(2):024102, July 2003. American Physical Society. DOI: 10.1103/PhysRevB.68.024102

  15. [23]

    M. I. Mendelev, D. J. Srolovitz, G. J. Ackland, and Sangcheol Han. Effect of Fe segregation on the migration of a non-symmetric Σ5 tilt grain boundary in Al. Journal of Materials Research , 20:208–218, 2005. DOI: 10.1557/JMR.2005.0024

  16. [24]

    C. G. Broyden. The convergence of a class of double-rank minimization algorithms 1. General considerations. IMA Journal of Applied Mathematics, 6(1):76–90, March 1970. DOI: 10.1093/imamat/6.1.76

  17. [25]

    Fletcher

    R. Fletcher. A new approach to variable metric algorithms. The Computer Journal, 13(3):317–322, 1970. DOI: 10.1093/comjnl/13.3.317

  18. [26]

    Goldfarb

    D. Goldfarb. A family of variable-metric methods derived by variational means. Mathematics of Computation, 24(109):23–23, 1970. DOI: 10.2307/2004873

  19. [27]

    D. F. Shanno. Conditioning of quasi-Newton methods for function minimization. Mathematics of Computation, 24(111):647–656, 1970. DOI: 10.1090/S0025-5718-1970-0274029-X. 23 A PREPRINT - F EBRUARY 13, 2025

  20. [28]

    M. R. Hestenes and E. Stiefel. Methods of conjugate gradients for solving linear systems. J. Res. Natl. Bur. Stand. (1934), 49(6):409, December 1952. National Institute of Standards and Technology (NIST). DOI: 10.6028/JRES.049.044

  21. [29]

    Sutton and Andrew G

    Richard S. Sutton and Andrew G. Barto. Reinforcement Learning: An Introduction. Cambridge University Press, Cambridge, MA, 2nd edition, 1986

  22. [30]

    Spinning Up in Deep Reinforcement Learning

    Joshua Achiam. Spinning Up in Deep Reinforcement Learning. https://spinningup.openai.com/en/ latest/, 2018

  23. [31]

    Reinforcement Learning: Industrial Applications of Intelligent Agents

    Phil Winder. Reinforcement Learning: Industrial Applications of Intelligent Agents . O’Reilly Media, Inc., Sebastopol, CA, 1st edition, 2021

  24. [32]

    Mailoa, Mordechai Kornbluth, Nicola Molinari, Tess E

    Simon Batzner, Albert Musaelian, Lixin Sun, Mario Geiger, Jonathan P. Mailoa, Mordechai Kornbluth, Nicola Molinari, Tess E. Smidt, and Boris Kozinsky. E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials. Nature Communications, 13(1), 2...

  25. [33]

    Euclidean neural networks: e3nn

    Mario Geiger, Tess Smidt, Alby M., Benjamin Kurt Miller, Wouter Boomsma, Bradley Dice, Kostiantyn Lapchevskyi, Maurice Weiler, Michał Tyszkiewicz, Simon Batzner, Dylan Madisetti, Martin Uhrin, Jes Frellsen, Nuri Jung, Sophia Sanborn, Mingjian Wen, Josh Rackers, Marcel Rød, and...

  26. [34]

    OpenAI Gym

    Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and Wojciech Zaremba. OpenAI Gym. arXiv preprint, 2016. arXiv:1606.01540

  27. [35]

    Chevrier, Kristin A

    Shyue Ping Ong, William Davidson Richards, Anubhav Jain, Geoffroy Hautier, Michael Kocher, Shreyas Cholia, Dan Gunter, Vincent L. Chevrier, Kristin A. Persson, and Gerbrand Ceder. Python Materials Genomics (pymatgen): A robust, open-source Python library for materials analysis...

  28. [36]

    Castelli, Rune Christensen, Marcin Dułak, Jesper Friis, Michael N

    Ask Hjorth Larsen, Jens Jørgen Mortensen, Jakob Blomqvist, Ivano E. Castelli, Rune Christensen, Marcin Dułak, Jesper Friis, Michael N. Groves, Bjørk Hammer, Cory Hargus, Eric D. Hermes, Paul C. Jennings, Peter Bjerre Jensen, James Kermode, John R. Kitchin, Esben Leonhard Kolsb...

  29. [37]

    Smidt, Mario Geiger, and Benjamin Kurt Miller

    Tess E. Smidt, Mario Geiger, and Benjamin Kurt Miller. Finding symmetry breaking order parameters with Euclidean neural networks. Physical Review Research, 3(1), January 2021. American Physical Society. DOI: 10.1103/physrevresearch.3.l012002

  30. [38]

    Glass, Artem R

    Colin W. Glass, Artem R. Oganov, and Nikolaus Hansen. USPEX—Evolutionary crystal structure prediction. Computer Physics Communications, 175(11):713–720, 2006. DOI: 10.1016/j.cpc.2006.07.020

  31. [39]

    Lillicrap, Jonathan J

    Timothy P. Lillicrap, Jonathan J. Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa, David Silver, and Daan Wierstra. Continuous control with deep reinforcement learning. arXiv preprint, 2019. arXiv:1509.02971

  32. [40]

    Understanding failures of deterministic actor-critic with continuous action spaces and sparse rewards

    Guillaume Matheron, Nicolas Perrin, and Olivier Sigaud. Understanding failures of deterministic actor-critic with continuous action spaces and sparse rewards. In Artificial Neural Networks and Machine Learning – ICANN 2020: 29th International Conference on Artificial Neural Ne...

  33. [41]

    Anubhav Jain, Shyue Ping Ong, Geoffroy Hautier, Wei Chen, William Davidson Richards, Stephen Dacek, Shreyas Cholia, Dan Gunter, David Skinner, Gerbrand Ceder, and Kristin A. Persson. Commentary: The Materials Project: A materials genome approach to accelerating materials innov...

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.