acceptodds
Under review as a conference paper at ICLR 2027

AtomWorld-Mirror: Macro-Step World Modeling of Critical Evolution Backbones for Materials Dynamics

Abstract

Atomistic simulation is a fundamental tool for studying long-term materials evolution, from diffusion and defect dynamics to interfacial reactions and fracture. Yet conventional simulators typically advance at microscopic resolution, spending substantial computation on low-impact local updates before reaching structurally consequential states—an evolutionary-resolution bottleneck that limits long-horizon simulation. We propose AtomWorld-Mirror, a time-aware macro-step world model for the critical evolution backbone of atomic systems. For Step-Wise atomistic simulation, AtomWorld-Mirror distills short micro-event segments into physically reachable transitions between key states, jointly predicting sparse structural edits and accumulated physical time through latent macro-step dynamics. Local reachability, inventory conservation, and continuous-time consistency constrain each transition. By amortizing local atomic physics into a reusable latent macro model and replacing explicit micro-event replay with macro-step inference, this formulation provides a path toward substantially faster prediction of long-term materials evolution while preserving structural validity and time semantics. Across five atomic systems, spanning Cu-rich RPV steel irradiation aging, Cu–Zr metallic glass, and LiN-based anti-perovskite solid electrolyte, macro-step inference delivers a speed up of to times over event-by-event simulation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
world models, atomistic simulation, AI for science

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_WbtKOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces a new method named AtomWorld-Mirror for simulating atomic dynamics. Instead of capturing fine-grained events at small time scales, the authors present a world model that operates at a macro scale, predicting atomic changes together with the physical time needed for those changes. The underlying model is described as an adaptation of DreamerV4.

Critical strengths
  • The method seems well aligned with the underlying dynamics problem considered in the paper.
Critical weaknesses

As a non-expert in atomic simulation, I assess the paper mainly from a general time-series modeling perspective. From this perspective, I have the following concerns:

  • What is the benefit of the latent space used here? In addition, why use a graph encoder rather than a CNN (or more particularly 3D CNN), given that the main problem operates on a regular lattice? The intuition behind these choices and the adaptations from DreamerV4 is insufficiently explained.

  • The overall framework looks quite complicated, with seven loss terms, each with its own weight. Figure 3 provides component ablations, but how sensitive is the method to these weights, and how were they chosen?

  • The comparisons mainly involve KMC variants and simplified versions of the authors’ own model. I am not convinced that these comparisons sufficiently justify the chosen learning model. Since the method's goal is to learn a stochastic dynamical system on a macro time scale, a diffusion model seems relevant. I believe a discussion of why the diffusion model is not considered here is necessary or better, a direct comparison on the same prediction task, including prediction quality and computational cost.

Questions for authors

See weaknesses.

9 comments
AuthorsOpenReview ↗

Response to Reviewer WbtK (1/4): latent space and graph encoder (W1)

Thank you for the review and for noting that the method is well aligned with the underlying dynamics problem. Below we explain the design choices behind the latent space and the graph encoder (W1) and the loss weighting (W2), and we report a direct comparison with a conditional diffusion model on the same prediction task, covering both quality and cost (W3). Since the submission we have continued to iterate on Mirror, and the new results will be included in the revised paper.

W1. Latent space, graph encoder, and adaptations from DreamerV4

Why a latent space. The latent space is where the teacher's physics lives inside Mirror. Training distills the atomistic laws encoded in the teacher simulator, namely which local events are possible, how fast they occur and how the energy changes, into Mirror's latent space, and aligns that space with the atomic physical world: the encoder maps a configuration to ztz_t, the macro dynamics predict where the system will be after kk events, and the latent and projection-consistency losses tie this prediction to the encoding of the true teacher endpoint (Eq. 23). Once the latent space is aligned with the physics, Mirror evolves the system directly in it, and there it learns and selects the transitions that truly matter, the key events of the evolution backbone, instead of replaying every low-impact micro event. This selection itself happens in latent space: the reward and duration heads score each candidate horizon from the latent states, and Eq. 24 picks the one with the largest predicted energy decrease per unit time. Three concrete consequences follow.

  1. Cost independent of system size. The macro dynamics act on a fixed-size global latent (32×1632\times16) plus the site embeddings of the active patch (Table 2), so one macro step costs about the same on any lattice. A model trained on 40340^3 runs on 2003200^3–4003400^3 without retraining at roughly 4–10 ms per macro transition (Sec. 5.4, Fig. 4).
  2. A place for the stochastic path. From the same state and horizon, different microscopic paths reach different endpoints in different times. The path latent hth_t captures this variation, and both the edit and the duration are conditioned on it (Eqs. 13–16, 20), so structure and time are predicted jointly.
  3. Measured benefit. An endpoint regressor that reuses the same graph encoder but predicts Xt+kX_{t+k} directly, without latent dynamics or a path latent, reaches change-F1 0.42, against 0.91 for Mirror (Table 6).

Why a graph encoder rather than a 3D CNN.

  1. The dynamics are sparse. At the vacancy density of 0.0002 (Table 1), a 40340^3 BCC lattice of 128,000 sites holds only about 26 vacancies, and every change happens in the candidate patches around them. The graph encoder processes the defect graph and the active patch, where events can occur. A dense 3D CNN would convolve every site at every step, so its cost grows with the lattice volume, which is exactly what a macro-step model must avoid at 4003400^3 and beyond.
  2. A BCC lattice is not a cubic grid. It consists of two interpenetrating sublattices, with 8 first- and 6 second-nearest neighbours. The KMC event neighbourhood (2NN hops, Table 1) maps directly onto graph edges, whereas a cubic convolution kernel would need a custom embedding of the lattice.
  3. The same encoder covers off-lattice systems. Cu–Zr metallic glass and the Li$_3$N-based electrolyte have no regular grid; the graph encoder handles them without change (Appendix D).
  4. Symmetry. Equivariant site features (Satorras et al., 2021) respect the lattice symmetries, consistent with KMC rates that depend only on the local environment.
AuthorsOpenReview ↗

Response to Reviewer WbtK (2/4): what is new relative to DreamerV4 (W1) and loss weights (W2)

What is new relative to DreamerV4. DreamerV4 contributes one general principle: learn a world model and roll it out in latent space. Mirror keeps this principle but changes what the world model models. DreamerV4 imitates an environment frame by frame at a fixed time step. Mirror is built around the critical evolution backbone (Sec. 3.1, Eq. 1): the long-term evolution of an atomic system is carried by a sparse sequence of key states, such as Cu-cluster growth and connectivity milestones in RPV steel or topology-changing cage swaps in the MD systems, separated by many low-impact micro events. Mirror predicts the next physically reachable backbone state together with the physical time needed to reach it, and skips the micro events in between. This backbone state–time target has no counterpart in DreamerV4, and every component is redesigned around it. These components are the contributions of Mirror:

ComponentDreamerV4Mirror (this work)What it enables
Supervisionrewards and observations from an environmentdistillation from a step-wise teacher simulator (KMC or MD) with continuous-time labels (Sec. 4.1)the latent space inherits the teacher's physical laws
State encodingimage framesgraph and equivariant encoder over the lattice, defect graph and active patch (Eq. 12)cost tied to the candidate patch; a 40340^3-trained model runs on 2003200^3–4003400^3
Transitionone fixed time stephorizon-conditioned macro transition to the next backbone state, with a path latent, kk up to 8192 (Eqs. 13–15)one step replaces 480 micro events on average (Table 7)
Outputreconstruction of the full observationsparse edit on the reachable support with inventory-conserving projection (Eqs. 16–19)zero reachability and inventory violations
Timestep index, no physical clockpath-conditioned expected and realized durations built on the CTMC estimate (Eqs. 20–22)cumulative clock ratio ρT=1.01\rho_T = 1.01 (Table 6)
Decisionaction policyhorizon selection in latent space: latent reward and duration heads score each horizon, Eq. 24 picks the largest energy decrease per unit timeeach macro step ends at the next key event

Together, these components turn a generic latent world model into a simulator-to-world-model distillation framework for atomistic state–time evolution. The key difference is that Mirror is designed as a distillation method rather than as a world model of one environment. Step-wise atomistic simulators that provide sparse structural states and physical-time supervision can expose the same teacher interface and be distilled into Mirror (Sec. 4), so the framework carries over readily to new systems and new teachers. DreamerV4 was only the starting point that inspired this design.

W2. Seven loss terms: sensitivity and choice of weights

The seven terms are not seven free knobs. The easiest way to read the loss is to ask what Mirror must output at each macro step. It has two main outputs that the simulation depends on: the structural change, and the time it takes. Every other term either helps the model learn these two outputs or keeps its latent variable well behaved, and the weights follow these roles (Table 1):

RoleTermWhat it teaches the modelWeight
Main outputsparse editwhich sites change, and into which species1.0
Main outputexpected duration τexp\tau_{\mathrm{exp}}how long the step takes on average1.0
Helperrewardthe energy drop of the step, used to rank horizons (Eq. 24)0.5
Helperlatentthe predicted next latent should match the encoding of the true teacher endpoint0.5
Helperprojection consistencythe state after the physical projection should match it too0.5
Helperrealized duration τreal\tau_{\mathrm{real}}the spread of the actual, random duration0.25
RegularizerKL; prior edit; prior latentthe rollout prior must imitate the training posterior, and the model is also trained in the mode it is used in0.05; 0.25; 0.25
AuthorsOpenReview ↗

Response to Reviewer WbtK (3/4): weight sensitivity (W2) and why not diffusion (W3)

The weights encode a simple priority. The two main outputs get the full weight of 1.0. Helpers get half weight, because they support the main outputs rather than being the goal. The realized duration gets 0.25 because its target is random by nature: even a perfect model cannot predict a single random draw, so a larger weight would mostly add noise to training. The KL weight is small (0.05), the usual choice for latent-variable models: large enough for the prior to learn from the posterior, small enough not to erase the path information carried by the latent.

Sensitivity. Fig. 3 already covers the most extreme setting of each weight. Across both Cu concentrations and all four temperatures, removing a term degrades the failure metric that term is responsible for (for example, removing the duration losses raises the time error), while the full model stays best. The weights therefore matter where they should, and the full configuration is not a fragile optimum: the same architecture and training loop carry over to KMC and MD teachers on five systems (Sec. 5.1, Appendix D).

W3. Diffusion models: discussion and direct comparison

Diffusion is in fact where Mirror started. DreamerV4, the latent world model that inspired Mirror, is itself a diffusion world model: its dynamics are trained with flow matching, and shortcut forcing reduces sampling to a few denoising steps. We deliberately replaced this diffusion dynamics with a single-pass, horizon-conditioned macro transition, because Mirror is built around one requirement: every macro step must be far cheaper than the micro events it replaces. The comparison the reviewer asks for is therefore also a comparison with the DreamerV4-style design, and it shows that this choice pays off in both quality and cost. The reasons are:

  1. Target. Mirror is built around the critical evolution backbone (Sec. 3.1, Eq. 1): each macro step goes directly to the next physically reachable key state, and the horizon is chosen in latent space so that the step ends at a key event (Eq. 24). This target is sparse and structured, since only a few sites along the backbone change. A diffusion model treats the endpoint as a generic high-dimensional sample and spends its denoising steps on every candidate site, including the many low-impact micro events that Mirror deliberately skips.
  2. Cost. Even with shortcut forcing, a diffusion world model needs several network passes per prediction, and Mirror must score every candidate horizon at every macro step. Mirror needs exactly one forward pass per horizon.
  3. Physical constraints. Mirror decodes a sparse edit on the reachable support Ck\mathcal{C}_k and projects it inside the loop. A diffusion model samples all candidate sites and can only be corrected afterwards.
  4. Physical time. Mirror's duration head is built on the CTMC estimate log⁡k−log⁡R(Xt)\log k - \log R(X_t) plus a learned residual (Eq. 20), so the rate dependence on temperature is carried by physics. A diffusion model has to learn the time scale from data alone, which generalizes poorly to unseen temperatures.
Reviewer_wZMwOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper tackles the problem of accelerating atomistic simulation by proposing AtomWorld-Mirror, a model that predicts both the atomistic state of the next meaningful evolution and the time required to reach it. The model is trained by distilling the microscopic simulation trajectories produced by conventional atomistic simulation methods. The proposed model is evaluated across five atomic systems, with a focus on speeding up the simulation processes of these systems, compared to the teacher (i.e., conventional atomistic simulation methods including MD and KMC).

Critical strengths
  1. The overall idea of the proposed method is interesting and has potential in accelerating simulation of atomistic systems.
  2. The design choices of the proposed method are sound and intuitive. Each design has its purpose and fits into the overall idea of the paper. The paper, for the most part, is easy to follow.
Critical weaknesses
  1. The primary weakness of the paper, as far as I can see, is the experimental setup and results. The experiments almost purely rely on comparison with the teacher methods (MD and KMC) the model is distilled from. The results convincingly demonstrate that the proposed method can accurately reproduce the simulation processes produced by the teacher methods while speeding up the processes significantly. However, more experiments and evaluation, especially those independent from the teacher methods, would improve the comprehensiveness of the experiments and provide more evidence of the usefulness of the proposed method. On the other hand, as discussed in the related work of the paper, there are several existing methods that share a similar idea of speeding up atomistic simulation with a learned model, but none of them are compared in the experiments.
  2. The paper does not have a discussion section about its limitations. The paper could discuss the utilization of conventional atomistic simulation beyond simulating evolution of atomistic systems, for example, computing thermodynamic and mechanical properties. Correct me if I am wrong, but the proposed method doesn't seem to have this kind of functionality.
Questions for authors
  1. The authors could debate on their rationale for only focusing on comparing the proposed method with the teacher methods in the experiments. Or better, provide more empirical evidence that the proposed method performs better than existing learning-based methods for accelerating atomistic simulation.
  2. The authors could discuss if the proposed method can be a drop-in replacement for conventional atomistic simulation methods in terms of various functionalities, and if not, where the limitation lies.
10 comments
AuthorsOpenReview ↗

Response to Reviewer wZMw (1/5): why the teacher is the reference (W1)

Thank you for the careful review, and for recognizing the idea and the design. We answer each weakness and question in turn.

W1. Evaluation beyond the teacher, and comparison with existing methods

Comparison with the teacher is the intended primary evaluation, and it follows from how the method is built. The concern suggests that the role of the teacher in our method may not have come through clearly in the paper, so we briefly restate the parts of the method that determine what the experiments must measure. AtomWorld-Mirror is a world-model distillation method. AtomWorld denotes the teacher, a step-wise atomistic simulator (KMC or MD) that supplies configurations, local event rules and the physical clock; Mirror is the macro-step world model distilled from it. In a distillation task the teacher is the ground truth, just as a learned interatomic potential is validated against the DFT data it is trained on.

The critical evolution backbone is defined on the teacher's trajectories. A step-wise simulator resolves every micro event at the same resolution, yet long-term evolution is carried by a small subset of structurally decisive states. In RPV steel aging, the Cu–vacancy exchanges that advance the Cu-cluster structure occur sparsely and at seed-dependent times within a fixed event budget (Fig. 1); most other events are transient fluctuations. The retained subsequence of key states, B=(Xn0,Xn1,…,XnM)\mathcal{B} = (X_{n_0}, X_{n_1}, \ldots, X_{n_M}) with M≪NM \ll N, is the critical evolution backbone (Eq. 1): Cu-cluster growth and connectivity milestones in RPV steel, and cage-swap topology changes in the MD systems. The macro horizon k=nm+1−nmk = n_{m+1} - n_m, up to 8192, is the number of teacher micro events between consecutive key states. The backbone is therefore a property of the teacher's dynamics, and it is a well-defined target: it persists across teacher seeds (persistence above 0.98, Sec. 4.4). Whether Mirror reaches the right key state can only be decided against this teacher backbone.

Mirror predicts the next key state directly, so what it must preserve are teacher quantities. From the current state XtX_t, Mirror jointly predicts the next backbone state and the physical time needed to reach it, pθ(ΔXt:t+k,τt:t+k∣Xt,Ck(Xt),k)p_\theta(\Delta X_{t:t+k}, \tau_{t:t+k} \mid X_t, \mathcal{C}_k(X_t), k) (Eq. 4), and selects kk adaptively so that each step ends at the next key state (Eq. 24). The structural change is a sparse edit restricted to the candidate set Ck(Xt)\mathcal{C}_k(X_t) reachable under the teacher's local event rules, and the duration is the expected CTMC time accumulated along the teacher path (Eq. 7). The low-impact intermediate events are never replayed; on average one macro step covers about 480 teacher micro events and 12.98 s of expected time (Appendix H). Skipping those events is only valid if the endpoint, the elapsed time and the energy change still match what the teacher would produce, which is exactly what edit F1, time log-MAE, the clock ratio and the energy error measure.

AtomWorld physics is distilled into the latent space, so every learned quantity is a teacher quantity. The teacher supplies segments (Xt,k,Xt+k,ΔX,τexp⁡,τreal,spath,Ck)(X_t, k, X_{t+k}, \Delta X, \tau_{\exp}, \tau_{\mathrm{real}}, s_{\mathrm{path}}, \mathcal{C}_k) (Eq. 9). A graph encoder maps the configuration to a latent state ztz_t (Eq. 12). A horizon-conditioned path latent hth_t captures how the teacher got from XtX_t to Xt+kX_{t+k}: during training it is inferred from the teacher path, and a KL term transfers this information into a prior that needs only the current state (Eqs. 13–14). The latent dynamics predict z^t+k=zt+Fθ(zt,ht,k)\hat z_{t+k} = z_t + F_\theta(z_t, h_t, k) (Eq. 15), from which the sparse edit, the duration and the energy change are decoded, under local reachability, inventory conservation and continuous-time consistency taken from the teacher's event rules. At rollout it evolves the system in its own latent space without calling the teacher, which is why we test the self-generated trajectory against fresh teacher probes started from Mirror's own states (Fig. 3).

Mirror changes the resolution at which the dynamics is advanced, not the physics. Its claim is that it reproduces the teacher's long-horizon evolution at a fraction of the teacher's cost, so agreement with the teacher is exactly what the experiments must measure.

AuthorsOpenReview ↗

Response to Reviewer wZMw (2/5): experiments beyond the teacher and a learned baseline (W1)

That said, the experiments are not limited to the teacher. Beyond per-segment agreement with one teacher run, the submission contains three kinds of experiments:

  1. Teacher run-to-run variability. The backbone is reproducible across teacher seeds: persistence above 0.98 for k=1,2,4,…,1024k = 1, 2, 4, \ldots, 1024 on 16 held-out trajectories, and largest-Cu-cluster identity 0.9066 across seeds (end of Sec. 4.4). A single teacher run is therefore not treated as exact truth.
  2. Conditions outside training. Fig. 3 runs Mirror closed loop on its own states and scores every transition against a fresh KMC probe, at Cu 0.005 and 0.0134 and 263–373 K, while training uses 300 K only (Table 1). A model trained on 40340^3 lattices is applied to 2003200^3–4003400^3 lattices (Fig. 4).
  3. Independent simulation software. Fig. 4 compares runtime with two accelerated-KMC variants, Rate Scaling KMC and Superbasin KMC, across lattice sizes and temperatures. Appendix I (Fig. 9) compares with OpenKMC (Li et al., 2019), a KMC code with extensive high-performance parallel optimization, on the 5.4×10105.4 \times 10^{10}-atom RPV steel system. OpenKMC runs on 5.2 million cores of the Sunway TaihuLight supercomputer, whose many-core processors are based on the SW26010 (Li et al., 2019), whereas AtomWorld-Mirror runs on a single server with eight NVIDIA A100 80 GB GPUs. As stated in the paper, that comparison is across different hardware and serves as a scale reference, not as a like-for-like ratio.

An independent learned model. In our response to Reviewer WbtK (W3) we also report a direct comparison with a learned generative alternative: a DreamerV4-style conditional diffusion Transformer that jointly samples the endpoint types on the candidate sites and the expected duration given (Xt,Ck(Xt),k)(X_t, \mathcal{C}_k(X_t), k). It is trained on the same teacher segments with the same encoder architecture (re-initialized), the same projection and the same training budget (120 epochs). Both models are evaluated on the same 24 states per condition, each with 12 independent teacher realizations, at 293 K and 353 K, Cu 0.0134 and 0.01, and kk from 4 to 1024; 353 K and Cu 0.01 were not seen in training.

MetricMirrorDiffusion (DreamerV4-style)
Distribution discrepancy to the teacher (0 = identical)0.003–0.0490.20–0.60
Median duration error, 293 K0.05–0.180.12–0.35
Median duration error, 353 K (unseen)0.02–0.10up to 4.9
Inference time per prediction12–15 ms158–280 ms

The discrepancy compares the set of endpoint edits sampled by each model with the teacher's realizations from the same state, and is zero when the two distributions coincide; the duration error is the median of ∣log⁡τ^−log⁡τexp⁡∣|\log \hat\tau - \log \tau_{\exp}|. This evaluation does not reuse the teacher runs Mirror was trained on: the reference is 12 fresh, independent teacher realizations per state, including conditions never seen in training, and the same reference judges both learned models. Here the teacher acts as a neutral referee between Mirror and an independent learned alternative, which is the comparison the reviewer asks for, and the comparison is clear on all three axes. Mirror is one to two orders of magnitude closer to the teacher distribution than the diffusion model (discrepancy 0.003–0.049 versus 0.20–0.60). It keeps the duration accurate at the unseen 353 K, where the diffusion model fails (median error 0.02–0.10 versus up to 4.9). It is about an order of magnitude faster per prediction (12–15 ms versus 158–280 ms, roughly 10–23×), and its cost stays flat from k=4k = 4 to 1024, whereas the diffusion model's cost grows with kk. A generic learned generative model trained on the same data with the same encoder therefore reaches neither Mirror's fidelity, nor its generalization to an unseen temperature, nor its speed.

AuthorsOpenReview ↗

Response to Reviewer wZMw (3/5): limitations, thermodynamic and mechanical properties (W2)

W2. Limitations; thermodynamic and mechanical properties

The paper does discuss its limitations. Appendix K ("Limitations and future work") states the scope of the evaluation, the current macro-step range k=1k = 1–81928192, and the extensions needed for larger species inventories and continuously evolving structures. Other limitations are stated where the corresponding results appear: the efficiency–coverage trade-off is reported explicitly (coverage 0.889 at k=128k = 128 and 0.649 at k=1024k = 1024, Table 7), and the cross-hardware nature of the OpenKMC comparison is stated in Sec. 5.4 and Appendix I. We agree that these points are scattered; this was a consequence of the limited space in the main text, and the camera-ready version will collect them into a Limitations section in the main text.

Thermodynamic and mechanical properties lie outside the task the paper sets, by design. The paper defines its task as long-horizon, large-scale atomistic evolution (title, abstract, Sec. 3): predicting the critical evolution backbone and the physical time needed to traverse it, in regimes where explicit event-by-event replay consumes most of the computational budget before structurally consequential states are reached (Appendix K). This is the function the teacher performs in our settings. KMC for RPV steel aging does not compute elastic constants or free energies either; it produces the time evolution of the configuration, such as Cu-rich precipitate formation under irradiation (Sec. 5.1). Mirror is designed to replace that function and is evaluated on exactly that function. Equilibrium thermodynamic sampling and mechanical testing are different uses of atomistic simulation, and the paper makes no claim about them.

Mirror's output is the input those calculations need. Mirror does not return a latent artifact but a physical configuration of the same lattice and species as the teacher: every predicted state is locally reachable and conserves species counts, with zero violations across all five systems (Sec. 5.1, Tables 3–6). Any property calculation applied to a teacher configuration therefore applies unchanged to a Mirror configuration. Mirror also carries physical quantities along the trajectory: the accumulated physical time (cumulative clock ratio 1.01, Table 6) and the energy change of each transition (the reward head, evaluated as the energy error in Fig. 3). In long-horizon studies the expensive part is reaching the evolved microstructure after a long physical time, and this is the part Mirror accelerates; the property evaluation on that microstructure is a single downstream calculation.

What Mirror does not do. It does not generate the microscopic path between key states, so short-time dynamics and time-correlation functions are not available. It does not sample equilibrium ensembles. Dynamics outside the teacher's event rules, including continuous off-lattice motion, are not represented. Each new system needs a small number of short teacher trajectory segments. The Limitations section of the camera-ready version will state these points explicitly; Q2 summarizes them.

Reviewer_4bTaOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
2
Soundness
3(+1)
Read review
Paper summary

AtomWorld-Mirror is a time-aware macro-step world model that distills step-wise atomistic simulators (KMC, MD) into a sparse-edit transition model for long-term materials evolution. The key insight is that atomistic simulators spend most of their time on low-impact local updates while structurally consequential events (precipitate growth, vacancy clustering, topology changes) occur sparsely. Mirror compresses a teacher KMC/MD trajectory into a sequence of macro transitions, each defined by an endpoint lattice edit, a latent state update, and an accumulated physical time. The model enforces three physical commitments -- local reachability, inventory conservation, and continuous-time consistency -- via reachability-constrained sparse edits and a duration head. Evaluated on five atomic systems: binary, 3‑element, and 10‑element RPV steel aging; Cu‑Zr metallic glass; Li3N‑based anti‑perovskite solid electrolyte, Mirror delivers 10^3-10^4x speedup over event-by-event simulation while preserving structural validity and temporal accuracy.

Critical strengths
  1. Novel cross-domain application of world models to materials science. Applying latent state-space rollout and sparse-edit prediction to atomistic simulation is a fresh and practically important direction. The "critical evolution backbone" framing -- identifying which structural states actually carry long-term materials evolution -- is a transferable insight beyond KMC/MD.
  2. Explicit physical constraints are baked into the model architecture. Reachability-constrained edits, inventory conservation via paired vacancy-atom selection, and continuous-time duration modeling are not post-hoc checks but are integrated into the training objective and inference procedure. This distinguishes the work from generic learned simulators that output unconstrained next-state predictions.
  3. Demonstrated speedup across multiple materials systems. The evaluation spans binary, 3‑element, and 10‑element RPV steel aging; Cu‑Zr metallic glass; Li3N‑based anti‑perovskite solid electrolyte -- covering alloy aging, metallic glass relaxation, and solid-electrolyte migration. The 10^3-10^4x end-to-end speedup with preserved structural validity (zero reachability and inventory violations across all five systems) and log-RMSE alignment is a strong practical result.
  4. Closed-loop validity testing. Rather than only evaluating single-step prediction accuracy, the paper runs long autonomous rollouts and scores reachability, inventory, and structural fidelity against the teacher trajectory. This is the right evaluation protocol for a macro-step model.
Critical weaknesses
  1. Generalization to out-of-distribution configurations is untested. Each individual system uses a fixed teacher configuration, and evaluation stays within the regime generated by that teacher. The paper does not show whether Mirror can interpolate to unseen compositions or temperatures, nor how it performs on completely different materials systems (e.g., oxides, ceramics). The claim of "10^3-10^4x speedup" is bounded to the narrow regime where the teacher distribution matches training.
  2. The speedup claim requires careful baseline interpretation. End‑to‑end wall‑clock timing including candidate construction, projection and neural inference is reported alongside OpenKMC runtime for the large‑scale benchmark. Speedup is defined as wall‑clock ratio, and skipping micro‑events is the core algorithmic mechanism of this macro‑step approach. Open questions remain regarding comparison against state‑of‑the‑art accelerated KMC baselines and detailed runtime decomposition of each model component.
  3. The model architecture is heavily specialized to lattice‑based KMC. While MD‑teacher systems (Cu‑Zr metallic glass, Li3N‑based electrolyte) are discussed and evaluated with snap‑to‑lattice bijection handling, continuous off‑lattice MD with arbitrary topology is not tested, which limits the generality of the "macro‑step world model" framing.
Questions for authors
  1. The 10^3-10^4x speedup is measured against event-by-event KMC/MD simulation. What is the baseline method -- are you comparing against a naive serial KMC, or against optimized KMC with parallel acceleration? How does the speedup change with system size (e.g., 40^3 lattice vs. the 5.4×10^10‑atom system)?
  2. The reachability-constrained edit set C_k(X_t) is derived from the teacher KMC at each step. During autonomous rollout, how does the model handle configurations that the teacher has not seen? Is there a coverage guarantee, or can the model drift into physically unreachable states?
  3. The duration prediction uses a log-space Gaussian NLL loss on tau_exp, with a lognormal auxiliary target for tau_real. Are these all within the training distribution, or does the evaluation include out-of-distribution temperatures or compositions?
  4. Inventory conservation (Eq. 18-19) is enforced as a hard constraint via paired vacancy-atom selection. How often does this constraint bind during rollout, and what is the failure rate when the model proposes an edit that violates conservation?
6 comments
AuthorsOpenReview ↗

Response to Reviewer 4bTa (1/4): generalization beyond the training regime (W1, Q3)

Thank you for the careful and constructive review, and for recognizing the critical-evolution-backbone framing, the built-in physical constraints and the closed-loop protocol. We address the weaknesses and questions below; W1 and Q3 are answered together, as are W2 and Q1.

W1 / Q3. Unseen temperatures and compositions, and other materials

The submission already evaluates Mirror outside its training regime, and we have added a direct interpolation test.

Already in the submission. The binary RPV model is trained at 300 K and Cu 0.0134 on a 40340^3 lattice (Table 1). The closed-loop ablation matrices (Fig. 3) evaluate it at 263, 293, 333 and 373 K and at Cu 0.005, none of which appears in training, and the timing diagnostic (Fig. 4) applies the same checkpoint to 2003200^3–4003400^3 lattices without retraining.

New: interpolation to unseen conditions. Following the reviewer's question, we trained Mirror on a grid of 263/293/333/373 K and Cu 0.005/0.0134 and evaluated it at 353 K and Cu 0.01, which lie between grid points and were never seen in training. For 24 states per condition, each with 12 independent teacher realizations and kk from 4 to 1024, we compare the endpoint edits sampled by Mirror with the teacher's realizations from the same state.

MetricValue
Distribution discrepancy to the teacher, all conditions (0 = identical)0.003–0.049
Median duration error ∣log⁡τ^−log⁡τexp⁡∣\lvert \log\hat\tau - \log\tau_{\exp}\rvert, 293 K (seen)0.05–0.18
Median duration error, 353 K (unseen)0.02–0.10

Accuracy at the unseen conditions is as good as at the seen ones. The duration is where interpolation would be expected to fail first, and the design explains why it does not: the duration head is the CTMC estimate log⁡k−log⁡R(Xt)\log k - \log R(X_t) plus a learned residual (Eq. 20), so the Arrhenius dependence of the rates on temperature and composition enters through R(Xt)R(X_t), computed from the current state, and the network learns only the path-dependent correction. A DreamerV4-style diffusion counterpart trained on the same data, which must learn the time scale from data alone, reaches a median duration error of up to 4.9 at 353 K (response to Reviewer WbtK, W3).

The duration evaluation is not limited to the training distribution. In the submission, the time-error column of the closed-loop matrices (Fig. 3) is measured at 263, 293, 333 and 373 K and at Cu 0.005, outside the 300 K / Cu 0.0134 training condition. In the new interpolation test, both duration targets are evaluated at the unseen 353 K and Cu 0.01: the τexp⁡\tau_{\exp} head reaches the median errors in the table above, and the lognormal τreal\tau_{\mathrm{real}} head gives 95% interval coverage of 0.92–0.94 (nominal 0.95).

Other materials systems. The submission already evaluates Mirror on materials of different classes, each with its own teacher (Appendix D): multi-component RPV steels with 3 and 10 elements (metallic alloys, KMC teacher), Cu–Zr metallic glass (an amorphous alloy, MD teacher) and the Li$_3$N-based anti-perovskite solid electrolyte (an ionic ceramic conductor, MD teacher). In every case the same architecture, constraints and training loop are distilled from that system's teacher; Mirror transfers as a method, and moving to a new system requires only a modest amount of teacher data, not new method design. We are actively extending the evaluation to further material classes and hope to report additional systems in the revised paper.

AuthorsOpenReview ↗

Response to Reviewer 4bTa (2/4): baselines, scaling and runtime decomposition (W2, Q1)

W2 / Q1. What the speed-up is measured against, and how it scales

Baselines. The reference in Fig. 4 and Appendix D is the teacher itself, an event-by-event, rejection-free residence-time (BKL) KMC with local rate updates, the standard optimized serial algorithm rather than a naive one. On the same hardware, Fig. 4 also compares against two accelerated KMC methods, Rate Scaling KMC (Lin et al., 2019), which down-weights fast event channels, and Superbasin KMC (Fichthorn & Lin, 2013), which aggregates recurrent low-barrier events, and reports Mirror's speed-up over each separately. Mirror remains faster than both, because it changes the prediction unit rather than the cost of sampling each event (Table 9).

Parallel KMC. Spatially parallel KMC, such as OpenKMC (Li et al., 2019), distributes event-by-event execution across domains; it reduces wall-clock time by adding hardware, not by reducing the number of events. It is complementary to Mirror rather than an alternative: Mirror reduces the number of steps, and its per-step cost is tied to local candidate patches, so it can also be distributed across domains. We therefore do not use OpenKMC for a like-for-like speed-up. Appendix I places the two at 5.4×10105.4 \times 10^{10} atoms only as a reference for the computing resources each route needs (5.2 million CPU cores versus one eight-GPU server).

Scaling with system size. Mirror's neural computation acts only on the candidate patches defined by the event neighbourhood, so its cost does not depend on the number of atoms: the batched forward pass stays at about 0.25 ms per macro step from 2003200^3 to 4003400^3 (table below), with a model trained on 40340^3 and no retraining. The only part of a macro step that grows with the lattice is reading the live state out of the simulator, and the end-to-end macro step stays at about 3–10 ms over this range (Fig. 4). The speed-up therefore does not shrink as the system grows; for k=1024k = 1024 it rises from about 1.1×1031.1 \times 10^3 at 2003200^3 to 2.3×1032.3 \times 10^3 at 4003400^3.

Where the speed-up comes from. We agree with the reviewer's reading: skipping low-impact micro events is the mechanism, so the speed-up is algorithmic, not a hardware effect. This is also why it is larger where each teacher event is more expensive, on the 10-element RPV steel (Fig. 6) and on MD teachers that integrate femtosecond steps between rare cage swaps (Fig. 7).

Runtime decomposition. We agree, and the measurements behind Fig. 4 already record it. Each macro step is timed end to end per trajectory as three components: the lattice readout, which turns the live state into model inputs (seed selection, candidate construction and ranking, defect-graph and patch features) and is charged once per state; the batched neural forward pass (path latent, latent dynamics, sparse-edit decoding, duration and reward heads); and the batched inventory-conserving projection. The last two run over an ensemble of 64 trajectories on one NVIDIA A100 GPU in bfloat16 and are divided by the ensemble width. For k=1024k = 1024 and Cu 0.005:

LatticeT (K)Readout (ms)Forward (ms)Projection (ms)Macro step (ms)Speed-up
2003200^32632.410.281.023.711092×
2003200^32932.480.270.503.251268×
2003200^33332.210.240.853.301251×
2003200^33732.410.270.593.261267×
3003300^32634.380.421.886.681497×
3003300^32934.590.301.396.271607×
3003300^33334.560.230.505.291955×
3003300^33735.050.231.166.431617×
4003400^32639.150.230.8810.262178×
4003400^32938.910.270.719.882250×
4003400^33339.210.270.8010.282266×
4003400^33739.050.270.7310.042302×

The neural forward pass is a small, constant share because it acts only on the candidate patches. The readout dominates, at 65–90% of the step, and it is the only component that grows with the lattice. We report this table in Appendix H, together with the hardware used.

AuthorsOpenReview ↗

Response to Reviewer 4bTa (3/4): candidate support at rollout, and how often inventory conservation binds (Q2, Q4)

Q2. Is the candidate set derived from the teacher during rollout? Can the model drift into unreachable states?

The candidate set is not taken from a teacher run. Ck(Xt)\mathcal{C}_k(X_t) is built from the current state and the horizon alone, by applying the teacher's local event rules (vacancy-mediated hops within the 2NN event neighbourhood, with the hop budget of Eq. 19) to XtX_t; Table 2 marks it as available at rollout time, and Algorithm 1 builds it inside the loop without calling the simulator. "Generated by the teacher dynamics" in Sec. 4.3 refers to these rules. The set is therefore defined for any configuration, including configurations the teacher has never visited.

This gives a guarantee of validity: every applied edit is reachable from the current state, so a rollout is a chain of reachable transitions and cannot enter an unreachable state. Zero reachability violations in all five systems follow from this construction. It does not guarantee fidelity: a reachable edit can still differ from what the teacher would most likely do, and such errors can accumulate. This is what the closed-loop protocol measures, by scoring each transition of the self-generated trajectory against a fresh teacher probe started from the model's own state (Sec. 5.3, Fig. 3). The interpolation results in our answer to W1 extend this test to unseen conditions.

Q4. How often does inventory conservation bind, and what happens when it fails?

Inventory conservation is not a fallback that triggers occasionally. Every predicted edit passes through the projection before it is applied (Algorithm 1, line 10), so the applied edit always conserves species counts and stays within the reachable set; a horizon whose edit cannot be made valid is discarded (line 12), and if none survives, the step stops with a recorded reason (line 14). The constraint therefore cannot be violated; what matters is how much the projection has to change the raw prediction. We measured this with the current version of Mirror in nine conditions (263–373 K, Cu 0.005 and 0.0134, 64 states each):

QuantityValue
Macro steps whose raw prediction changes a species count (projection disabled)15–64%
Atoms or vacancies claiming an already-occupied site per step (projection disabled)0.17–1.67
Reachability and inventory violations with projection0% in all conditions
Movers per prediction that no sampled target can place, k≤1024k \le 10240.01–0.57 on average (maximum 1.27)

The projection is active in a large fraction of steps but changes few sites, and it leaves the distribution-level agreement with the teacher unchanged (discrepancy 0.0005–0.006 with projection, against 0.0001–0.006 without it).

Reviewer_tYUiOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes AtomWorld-Mirror, a teacher student framework for accelerating atomistic simulation through learned macro step transitions. Using trajectories from conventional simulators, the model jointly predicts sparse structural edits and accumulated physical duration. The framework combines a path-conditioned latent model, candidate restricted editing, inventory preserving projection, and adaptive horizon selection to generate successive macro states. Experiments cover five material systems with KMC and MD teachers, evaluating structural agreement, timing accuracy, constraint satisfaction, and computational efficiency, and report 10e3 to10e4 speedups over event by event simulation.

Critical strengths
  1. The problem is well motivated, and appropriate. Compressing microscopic events into macro transitions addresses an important computational bottleneck in materials simulation. Explicitly accounting for path dependent duration makes the formulation more relevant to physical evolution than endpoint prediction alone.

  2. The architecture incorporates useful domain structure and shows promising local prediction results. Sparse editing, candidate restrictions, and inventory preserving projection are appropriate inductive biases for the primary atom-vacancy setting. Fig. 2 demonstrates encouraging structural and temporal agreement with the teacher, while fig.3 provides evidence that the proposed components contribute to the measured performance.

Critical weaknesses
  1. The evidence does not sufficiently establish the physical and statistical fidelity of the proposed method. The main reported metrics such as edit F1, pointwise duration errors, and constraint-violation rates provide useful evidence of local prediction quality, but as far as I can tell does not sufficiently demonstrate the preservation of the relevant transition statistics. In particular, edit F1 requires careful interpretation when the reference trajectory is stochastic, while zero violations primarily demonstrate that the imposed support and inventory constraints are enforced. Sec 5.2 mentions duration NLL, probability-integral-transform diagnostics, and interval coverage, but I was unable to locate the corresponding results in the main text or the appendix.

  2. The efficiency results would benefit from clearer workload and timing definitions. Fig 4 and 6 show milliseconds scale Mirror runtimes, whereas Table 7 reports a throughput of 0.713 segments per second. Fig 9 presents second scale Mirror runtimes for a benchmark described as covering 54 billion atoms and 50 years of physical evolution. These measurements may well correspond to different workloads, but I could not find an explanation of how they relate to one another. Sec 5.4 lists the operations included in the timing. Even so, it remains unclear to me which results measure individual macro transitions, which measure batched diagnostics, and which measure complete autonomous trajectories. The implementation and hardware of the large scale OpenKMC benchmark should also be distinguished from those used in the other timing comparisons. Since the acceleration factor is a headline claim, clear benchmark specific reporting seems necessary to assess it.

Questions for authors

My background is primarily mathematical rather than in materials simulation, so some of the questions below may due to my own misunderstanding of the domain. I would be grateful for the authors’ clarification.

  1. Is Mirror intended to approximate the joint conditional distribution of endpoint structure and duration, or only particular statistics of that distribution? When computing edit F1, how are model and teacher realizations paired, and are the model predictions generated exclusively from the rollout prior?

  2. Could the authors specify the rule used to identify key states along a teacher trajectory and define the reward in eq.24? I may be missing something, but it is unclear to me how selecting among randomly generated candidates by maximizing predicted progress per unit time relates to the transition statistics that Mirror is intended to reproduce.

7 comments
AuthorsOpenReview ↗

Response to Reviewer tYUi (1/4): Q1

Thank you for the careful reading and for the constructive questions. We answer the two questions first, because they define what the evidence should be measured against, and then address the two weaknesses. We also note that, since the submission, we have continued to iterate on the model. The updated version agrees better with the teacher on several of the diagnostics requested here, in particular the distribution-level comparison and the calibration of predicted durations, and we will include these new results in the revised paper.

Q1. What does Mirror approximate, and how is edit F1 paired?

Target. Mirror is designed to approximate the joint conditional of Eq. 4, and this is the point of the framework. Structure and time are generated together: the path latent hth_t comes from the rollout prior (Eq. 14), and the edit decoder, the expected-duration head and the lognormal head for the realized duration τreal\tau_{\mathrm{real}} are all conditioned on the same hth_t (Eqs. 15, 16, 20–22). The coupling between which endpoint is reached and how long it takes is therefore part of the model. In the submission we evaluate this joint model through the statistics that govern long-horizon evolution, which is how accelerated KMC methods are usually validated and which needs only one teacher realization per state: (i) the projected sparse edit on Ck(Xt)\mathcal{C}_k(X_t); (ii) the expected clock τexp\tau_{\mathrm{exp}} (Eq. 7), whose sum along a trajectory is checked by the cumulative clock ratio (ρT=1.01\rho_T = 1.01, Table 6); and (iii) the accumulated energy change. These statistics are well defined under teacher stochasticity, because the backbone is reproducible across seeds: persistence above 0.98 for k∈1,2,4,…,1024k \in \\{1, 2, 4, \ldots, 1024\\} on 16 held-out trajectories, and largest-Cu-cluster identity 0.9066 (end of Sec. 4.4). The distribution-level comparison under W1 tests the joint model directly, and the updated Mirror shows improved agreement on it. We will make this explicit in Sec. 3.2.

Pairing. Mirror uses Multi-K: at each step it predicts candidates over k∈1,…,1024k \in \\{1, \ldots, 1024\\} (our current version of the model extends this to 8192, which will be updated in the paper) and selects the horizon adaptively, so that each macro step ends at the next key event (Eq. 24). For edit F1, each segment pairs one model prediction with one teacher realization: both start from the same state XtX_t and cover the same selected horizon k⋆k^\star. The model side is Mirror's projected edit for the selected k⋆k^\star, i.e. the Multi-K candidate it chose, computed with the path latent from the rollout prior; the teacher side is the change mask mi=1xt+k,i≠xt,im_i = \mathbb{1}\\{x_{t+k,i} \neq x_{t,i}\\} of that run (Eq. 17). F1 is the harmonic mean of precision and recall between the two change masks over the candidate support Ck(Xt)\mathcal{C}_k(X_t) (Sec. 5.2), type accuracy is reported separately, and both are averaged over segments. The two protocols differ only in where XtX_t comes from:

  • Teacher-forced (Fig. 2, Tables 3–6; Appendix F): from the teacher state XtmX_{t_m}, Mirror selects k⋆k^\star and predicts the projected edit; the teacher then advances the matching k⋆k^\star micro events, and the edit is compared with the change mask of that realization over Ck\mathcal{C}_k.
  • Closed loop (Fig. 3, Sec. 5.3): the model carries its own state, and each selected transition is scored against a KMC probe started from that model state.

Prior only. Yes. Every evaluated prediction comes from the rollout prior pθ(ht∣zt,gt,k)p_\theta(h_t \mid z_t, g_t, k); the posterior qϕq_\phi needs the teacher endpoint and path summary and is training-only (Table 2; Fig. 5). For each (Xt,k)(X_t, k) we take the transition given by the prior mean. A rollout must commit to one state per macro step (Algorithm 1, line 16), and Eq. 24 compares one candidate per horizon, so the evaluation uses the representative transition of the learned distribution. This also keeps runs reproducible and matches the per-transition timing of Figs. 4 and 6.

AuthorsOpenReview ↗

Response to Reviewer tYUi (2/4): Q2

Q2. Key-state rule, the reward in Eq. 24, and transition statistics

Key states. For RPV steel (KMC teacher), a teacher segment ends at a Cu–vacancy exchange, the events marked in Fig. 1, and is capped at k=1024k = 1024 (our current version of the model extends this to 8192, which will be updated in the paper). For Cu–Zr and Li$_3$N-based systems (MD teacher), a macro step ends at a cage-swap topology event (Appendix D). We will state these as checkable predicates in Sec. 3.1.

Reward. The reward of a segment is its energy decrease, r=10.0×(Ebefore−Eafter)r = 10.0 \times (E_{\mathrm{before}} - E_{\mathrm{after}}), where 10.0 is a fixed scale (Sec. 5.2, Table 2) that does not change any ranking. A positive reward means the system has moved to a more stable configuration; in RPV steel this mainly reflects Cu clustering, which is the precipitation process the simulation is meant to capture. Eq. 24 divides the predicted energy decrease by the predicted duration, so among the legal horizons it picks the one with the fastest structural progress per unit physical time. As an illustration, if k=16k = 16 is predicted to lower the energy by 0.2 eV in 1 s and k=256k = 256 by 1.0 eV in 20 s, Eq. 24 picks k=16k = 16 (0.20 vs. 0.05 eV/s): the longer step spends most of its time on low-impact fluctuations. Because the energy drops sharply at key events, this rule ends each macro step at the next key event, which is how Multi-K adapts the horizon to the distance to that event.

The candidates are not randomly generated. This point is central to the method, so we state it precisely. At each macro step, Mirror encodes XtX_t once and then, for every horizon kk in the Multi-K set, computes exactly one candidate (Algorithm 1, lines 5–13): it builds the candidate support Ck(Xt)\mathcal{C}_k(X_t), takes the path latent from the rollout prior, predicts the next latent, decodes the sparse edit, projects it to an inventory-conserving edit, and predicts its energy decrease and duration. At evaluation the path latent is the prior mean, so each candidate is fully determined by (Xt,k)(X_t, k); there are no repeated random draws at the same kk and no sampling-and-filtering. The result is one table with one row per legal horizon, and Eq. 24 simply picks the row with the largest predicted energy decrease per unit time. The candidates for all horizons are computed in one batched pass, which is also why one macro transition costs only milliseconds (Figs. 4 and 6).

Relation to transition statistics. Eq. 24 never creates or modifies a transition. Each row is produced by the edit and duration heads, which are trained on teacher segments at every kk (Eqs. 17, 21–23), so the transition statistics live in those learned heads. Eq. 24 only decides where the trajectory stops, much like adaptive step-size control in a numerical integrator, which chooses the step length without changing the equations being solved. Whichever k⋆k^\star is chosen, the clock advances by the duration predicted for that same k⋆k^\star, so structure and time stay consistent.

Evidence. The data confirm this directly. Under the Eq. 24 planner the cumulative clock ratio is ρT=1.01\rho_T = 1.01 (Table 6), whereas the unconstrained variant reaches only 0.35; over about 200 consecutive macro segments the final edit ratio is ≈0.92\approx 0.92 and the final clock ratio ≈1.05\approx 1.05 (Fig. 2a, 2c). Any systematic bias introduced by the selection, toward faster or more energy-favourable transitions, would accumulate along these trajectories and appear as drift in exactly these quantities. No such drift is observed. Horizon selection therefore leaves both the clock and the structural progress intact.

AuthorsOpenReview ↗

Response to Reviewer tYUi (3/4): W1

W1. Physical and statistical fidelity

The submission evaluates fidelity well beyond local, per-segment metrics, at three levels:

  1. Trajectory level (Fig. 2). Over about 200 consecutive macro segments, cumulative edits (final ratio ≈0.92\approx 0.92), cumulative expected time (final ratio ≈1.05\approx 1.05) and structurally faithful steps (95.1%) track the teacher. These are accumulated quantities: any bias in the transition statistics would compound along the trajectory and appear as drift, and none does.
  2. Closed loop against fresh KMC probes (Fig. 3). The model runs on its own states, and every transition is scored against a KMC probe started from that state, at two Cu concentrations and four temperatures (263–373 K), while Table 1 lists a single training temperature of 300 K.
  3. Discriminative baselines (Table 6). On the same metrics, the unconstrained Dreamer-style variant obtains log-MAE 4.08, scale 0.35 and ρT=0.35\rho_T = 0.35, and endpoint regression obtains change-F1 0.42, against 0.13, 1.00, 1.01 and F1 0.91 for Mirror. The metrics are therefore far from trivially satisfied.

Edit F1 under a stochastic reference. As detailed in Q1, F1 is computed at identical state and horizon, on a backbone that is reproducible across teacher seeds (persistence above 0.98, cluster identity 0.9066).

Zero violations. This is a guarantee by design and one of the contributions, not a by-product. Local reachability and inventory conservation are physical requirements: a step that violates them produces a state the teacher could never reach. Learned models do not satisfy them on their own: without the constraints, 54.8% of edits are unreachable and 40.8% change the species counts (Table 6). Enforcing them by construction is what keeps long autonomous rollouts physical.

Duration NLL, PIT and coverage. These diagnostics concern the auxiliary τreal\tau_{\mathrm{real}} head; the primary time target τexp\tau_{\mathrm{exp}} is evaluated throughout the paper by log-MAE, scale ratio and ρT\rho_T. We will report the τreal\tau_{\mathrm{real}} diagnostics per system in the revision, together with the updated version of Mirror, which is well calibrated on them.

To make this evidence even more direct, the revision adds:

  1. Duration NLL, PIT and 68%/95% coverage for τreal\tau_{\mathrm{real}}, the diagnostics announced in Sec. 5.2.
  2. A distribution-level comparison over repeated teacher seeds from the same XtX_t: duration, energy change and Cu-cluster statistics, with Wasserstein or KS distances.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Looking forward to feedback on the Responses

Hello! Since the discussion period has started, we would like to kindly ask you to take a look at our responses and reevaluate our work based on our clarifications. Please let us know whether our response addresses your concerns or whether there is any further detail we can provide to help address them. We appreciate your time and consideration!

The Authors

AuthorsOpenReview ↗

Summarization on the Responses

Thanks for handling and reviewing our submission "AtomWorld-Mirror: Macro-Step World Modeling of Critical Evolution Backbones for Materials Dynamics". We would like to thank all reviewers for their insightful and constructive comments. The reviewers found the idea novel and well motivated (tYUi, 4bTa, wZMw), the physical constraints built into the model sound (4bTa, wZMw, tYUi), and the closed-loop validation and multi-system speed-up convincing (4bTa). Addressing each point has substantially strengthened the paper, and we added several new experiments during the discussion. The major responses are summarized as follows:

(1). We evaluated Mirror at the level of the full joint distribution of structural change and physical duration, on 128 held-out states with stochastic samples from the rollout prior. Mirror matches the teacher's marginals at or close to the teacher's own sampling floor, reproduces the horizon-dependent coupling between energy change and duration, and gives calibrated duration intervals (coverage 0.67–0.71 and 0.94–0.97 for the 68% and 95% intervals) (See Reviewer tYUi.Q1, tYUi.W1 and the distribution-level follow-up).

(2). We clarified the key-state rule, the reward in Eq. 24, and the workload and timing definition of every speed-up result in one table (See Reviewer tYUi.Q2, tYUi.W2).

(3). We showed generalization beyond the training regime: Mirror trained on a temperature–composition grid matches the teacher at an unseen temperature and composition as well as at seen ones (discrepancy 0.003–0.049, median duration error 0.02–0.10 at the unseen 353 K) (See Reviewer 4bTa.W1&Q3).

(4). We added a new material class, oxygen-ion conduction in yttria-stabilized zirconia, an oxide with charged defects and long-range electrostatics. Mirror matches the teacher at its sampling floor at a held-out temperature and composition, reproduces the activation energy (0.624 / 0.624 eV) and the non-monotonic dependence of diffusivity on dopant content, with zero constraint violations and directly measured speed-ups of 761× to 9,437× (See Reviewer 4bTa follow-up, W1).

(5). We extended Mirror to continuous off-lattice reactive MD with no reference structure, ReaxFF pyrolysis of n-octane, where molecules form and break. Mirror reproduces the decomposition of octane and the formation of ethylene, H2 and methane, with zero valence violations; the network alone is 377–1,047× faster than MD (See Reviewer 4bTa follow-up, W3).

(6). We clarified the baselines and scaling of the speed-up (optimized serial BKL KMC, Rate Scaling KMC and Superbasin KMC), and showed that the candidate set is built at rollout without the teacher, so every rollout step is reachable and inventory-conserving by construction (0% violations in all conditions) (See Reviewer 4bTa.W2&Q1, 4bTa.Q2&Q4).

(7). We compared Mirror with the published learned acceleration method for the same application, ANN-based KMC (Messina, Castin, Domain and Olsson, Phys. Rev. B 2017), the standard learned approach for RPV steel aging. Both reproduce the Cu precipitate structure within about two standard deviations of an exact KMC reference, and Mirror reaches the same physical time 134–177× faster; the gap widens with system size and simulated time (See Reviewer wZMw follow-up).

(8). We compared Mirror with a DreamerV4-style conditional diffusion Transformer trained on the same data with the same encoder. Mirror is one to two orders of magnitude closer to the teacher distribution, stays accurate at an unseen temperature where the diffusion model fails, and is about an order of magnitude faster per prediction (See Reviewer WbtK.W3, wZMw.W1&Q1).

(9). We explained the role of the latent space and the graph encoder, what is new relative to DreamerV4, and the loss weights and their sensitivity (See Reviewer WbtK.W1&W2).

(10). We explained why agreement with the teacher is the right primary evaluation for a distillation method, stated the limitations, and specified where Mirror can and cannot replace a conventional simulator (See Reviewer wZMw.W1, W2, Q1&Q2).

Again, we would like to sincerely thank the AC and all reviewers for the constructive comments and the time spent on our manuscript.

Best regards,

The Authors of Submission 16268

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.