{"id":"32872b00-852b-460d-91ec-f234b42bd2ad","arxiv_id":"2502.08405","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A deep reinforcement learning agent can relax some Al-Fe crystals in fewer steps than BFGS or conjugate gradient, but the main benchmark is biased because the agent is trained on the test structures.","lead":"Researchers trained a deep reinforcement learning agent to relax aluminum-iron crystal structures, testing it on model interatomic potentials. The agent matches or beats classical optimizers in some settings, but only after being trained on the very structures it is later asked to relax.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No wall-clock or per-step cost is reported; the step-count advantage in §3.3 cannot substantiate 'acceleration' if a TFN policy evaluation is more expensive than an EAM force call.","rationale":"The reader's weakest_assumption is exactly that the paper equates fewer relaxation steps with acceleration without reporting wall-clock or per-step cost. My analysis independently identifies this as the load-bearing flaw: it directly invalidates the title and abstract claim, not merely the scope of the method. The memorization issue raised in the reader's rationale is also real, but the paper's conclusion explicitly scopes the method to repeated relaxation of identical or closely related structures, so for that use case training on the target structure is legitimate. The cost comparison, by contrast, is required even for the narrow claim. The authored text contains no timing information anywhere and only step counts in Fig. 3c, Fig. 5a, and Table 1, so the concern is grounded in the manuscript's own evidence. A proper fix would be to add a benchmark with wall-clock time or per-step cost; this is a concrete, testable check. Because the reader already rejected the paper on essentially this basis, my verdict remains unchanged. I do not find a separate fatal flaw in the methodology; the RL framework, reward design, and generalization experiments are presented honestly, including the low success fractions of 0.1–0.4 in Section 3.4.3. The paper could become a valuable contribution with a corrected cost analysis and, ideally, an unbiased test on genuinely unseen structures.","tokens_in":21234,"tokens_out":2855,"duration_ms":34823,"concrete_test":"Reproduce the Section 3.3 benchmark for N=10 and N=20 using the published code, and measure wall-clock time per complete relaxation run (or average time per step) for the trained TD3 agent versus BFGS and conjugate gradient on the same hardware and with the same force thresholds. If the TFN policy evaluation plus environment update does not make the RL agent's total wall-clock time lower than BFGS/CG, the step-count advantage does not support the 'acceleration' claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim, stated in the abstract and title, is that DRL models can 'enhance the efficiency of structure relaxation' or 'accelerate' it compared to classical optimizers. The only quantitative support is Section 3.3 (Fig. 5a), where the trained TD3 agent needs fewer relaxation steps than BFGS or conjugate gradient for N=10 and N=20 structures at force thresholds of 0.2 and 0.25 eV/A. The paper never reports wall-clock time, CPU/GPU time, or per-step cost. This matters because the two methods have different per-step costs. In the RL rollout, each step requires a TFN policy evaluation (a multi-layer E(3)-equivariant graph network forward pass) plus a pymatgen/ASE environment update with an EAM force evaluation. In a classical BFGS or CG relaxation, each step requires one or more EAM force evaluations plus a low-cost numerical update. Since EAM potentials are very cheap for N=10–20 atoms, the extra TFN inference could easily dominate the per-step cost. If that is the case, a 2–3x reduction in steps would not translate into faster relaxation. The conclusion's own suggestion to switch to classical optimizers below 0.1–0.01 eV/A further underscores that the claimed advantage is confined to a moderate-force regime, but the missing cost comparison is the decisive gap: without it, 'acceleration' is an unsupported inference even within that regime. The paper is otherwise technically transparent and provides code and data, but the central claim as stated is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":21464,"tokens_out":6584,"duration_ms":62735,"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":[{"comment":"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.","section":"Section 3.3, Fig. 5a"},{"comment":"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.","section":"Section 3.3"},{"comment":"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.","section":"Section 3.3 / Fig. 5a caption"},{"comment":"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.","section":"Section 3.4.3, Fig. 7"}],"minor_comments":[{"comment":"Typo in the subsection heading: 'Tensor-Filed' should be 'Tensor Field'.","section":"Section 2.3"},{"comment":"Typos: 'demostrate' should be 'demonstrate', and 'exhausting experience' should be 'exhaustive experience'.","section":"Appendix C"},{"comment":"Typo in the caption: 'relaxion' should be 'relaxation'.","section":"Appendix D, Fig. 8 caption"},{"comment":"The expression for the geometric sum is ambiguous as typeset; please write it as (1−γ^n)/(γ−1) to avoid confusion.","section":"Eq. (13)"},{"comment":"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.","section":"Section 3.4.3"},{"comment":"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).","section":"Appendix K"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is honest and provides code/data, but the title and abstract overstate the results. The two main issues—the missing wall-clock comparison and the trained-on-test benchmark—are fixable in revision. I would not reject the paper on the basis of novelty or field fit; the equivariant-RL approach to relaxation is interesting. The main question is whether the authors can supply the missing cost measurements and a fair baseline. If not, the claims should be substantially weakened."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a transparent, well-documented proof-of-concept for RL-based crystal relaxation, and it is not the universal optimizer the title hints at. The 'acceleration' claim is not actually established.\n\nWhat's new: applying TD3 with an E(3)-equivariant TFN to relaxation, with exploration tricks (noise to state, greedy exploration) that respect crystal symmetry. The authors compare architectures and rewards carefully, and they are honest about failures: CGCNN doesn't work, step reward causes convergence problems, SAC underperforms. They share code and data, and the hyperparameter study is unusually thorough.\n\nSoft spots in proportion: the primary benchmark (Fig. 5a) trains on the exact structure it tests on, so the step-count advantage over BFGS/CG is largely memorization. The generalization experiment (Fig. 7) shows success fractions of 0.1–0.4 on unseen structures, which they admit is insufficient. And no wall-clock or per-step cost is reported; since EAM is cheap and a TFN forward pass is not, step counts alone don't justify 'acceleration.' The force plateau at 0.1–0.2 eV/A further limits the practical claim. These are not fatal to the paper as a proof-of-concept, but they are decisive against the title and abstract wording.\n\nThe paper is still useful. It identifies real design issues (symmetry-preserving exploration, reward shaping, sensitivity near minima) and gives a clear picture of what's needed next. The claims should be toned down: 'potential' is fine, 'acceleration' is not supported.\n\nThis deserves a serious referee and likely major revision. If the authors redo the benchmark on unseen structures and report wall-clock or per-step costs, it could be a solid contribution. I'd engage with it rather than desk-reject.","headline":"Careful proof-of-concept for RL relaxation, but the 'acceleration' claim rests on memorized benchmarks and missing wall-clock costs.","tokens_in":22094,"tokens_out":1940,"would_cite":false,"duration_ms":19936,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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.","keywords":["structure relaxation","reinforcement learning","E(3)-equivariant neural networks","tensor field networks","TD3","crystal graphs","Al-Fe alloys","embedded-atom method"],"falsifier":"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.","tokens_in":20925,"feed_emoji":"⚛️","tokens_out":8178,"duration_ms":84823,"temperature":0.7,"pith_summary":"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.","feed_headline":"Fewer steps to relax crystals: RL beats BFGS","feed_subtitle":"A TD3 agent trained on Al-Fe structures reaches moderate force thresholds in fewer steps and transfers to new compositions.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Supplies the crystal-graph representation of structures and the CGCNN architecture that serves as the baseline that fails on this task.","marker":"[17]"},{"why":"Supplies the E(3)-equivariant tensor field network architecture that carries the agent's symmetry-preserving policy.","marker":"[18]"},{"why":"Supplies the TD3 actor-critic algorithm used for all main training runs.","marker":"[19]"},{"why":"Supplies the SAC algorithm whose comparison with TD3 motivates the choice of TD3 for structure relaxation.","marker":"[20]"},{"why":"Defines the EAM potential for Fe that generates the potential energy surface for Fe-containing structures.","marker":"[21]"},{"why":"Defines the EAM potential for Al used for Al-containing structures in the training data.","marker":"[22]"},{"why":"Defines the EAM potential for Al-Fe used for the binary compounds and the larger N=8, 10, 20 test cells.","marker":"[23]"},{"why":"Provides the Curie-principle argument that an E(3)-equivariant model's output symmetry can only match or exceed the input symmetry, justifying the restricted action space.","marker":"[37]"},{"why":"Generates the random Al-Fe structures used for the N=8, 10, 20 benchmarks and the transfer datasets.","marker":"[38]"},{"why":"Explains why deterministic actor-critic methods stall on piecewise-constant reward functions, supporting the paper's step-reward analysis and reward-switching procedure.","marker":"[40]"}],"fun_headline_variants":["RL relaxes crystals in fewer steps than BFGS","Symmetry-aware DRL cuts steps to relax crystals","Equivariant RL beats BFGS for crystal relaxation","Deep RL learns to relax crystals efficiently","Symmetry-enforced RL speeds crystal relaxation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["RL relaxes crystals in fewer steps than BFGS","Symmetry-aware DRL cuts steps to relax crystals","Equivariant RL beats BFGS for crystal relaxation","Deep RL learns to relax crystals efficiently","Symmetry-enforced RL speeds crystal relaxation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001166,"raw_usage":{"total_tokens":4812,"prompt_tokens":919,"completion_tokens":3893,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":535,"completion_tokens_details":{"reasoning_tokens":3822}},"tokens_in":535,"tokens_out":3893,"duration_ms":35542,"temperature":1.0,"reasoning_tokens":3822,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T05:12:50.543528+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the EAM potential for Fe that generates the potential energy surface for Fe-containing structures."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the EAM potential for Al-Fe used for the binary compounds and the larger N=8, 10, 20 test cells."},{"cited_title":"Understanding failures of deterministic actor-critic with continuous action spaces and sparse rewards","cited_arxiv_id":null,"evidence_quote":"Explains why deterministic actor-critic methods stall on piecewise-constant reward functions, supporting the paper's step-reward analysis and reward-switching procedure."}],"review_version":1}