acceptodds
Under review as a conference paper at ICLR 2027

FLARE: Flow Matching with Local Axis-Angle Representations for Stochastic Micromagnetic Evolution

43 views

Abstract

Micromagnetic dynamics pose a structured stochastic generative problem: high-dimensional fields of unit-vector magnetizations can evolve into multiple possible futures under the same physical controls. Existing learned approaches generally retain stepwise integration or model deterministic evolution, leaving full-field, direct-horizon stochastic prediction largely unexplored. We propose FLARE, a flow-matching framework that recasts stochastic finite-time magnetization prediction as conditional transport over anchor-relative local axis-angle rotations. This rotation-space formulation respects the intrinsic geometry of magnetization dynamics and preserves pointwise unit norm by construction. By explicitly conditioning on the physical prediction horizon, FLARE directly generates full-field stochastic endpoints across multiple target times without stepwise integration. Across seven author-released learned baselines, FLARE leads all three distributional metrics, including 29.9% lower angular energy distance and 37.3% lower fair energy score than the strongest external baseline on each metric. On a representative composed 5-ns protocol, it further achieves a best-batch speedup over MuMax on a single GPU.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Stochastic micromagnetics, Generative modeling, Ensemble-scale simulation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_XZx7OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper applies conditional flow matching to stochastic micromagnetic prediction using anchor-relative axis–angle representations. The approach preserves unit magnetization norm and generates future states directly, improving distributional accuracy over evaluated baselines while accelerating prediction relative to numerical simulation.

Critical strengths

s1) The application for the tested framework is interesting and have not been tested before.

Critical weaknesses

w1) The method combines standard conditional flow matching with established axis angle coordinates. The methodological novelty beyond adapting these components to micromagnetics is limited.

w2) The experiments demonstrate effectiveness within one physical system, but do not establish a generalizable methodological insight or a substantial scientific advance enabled by the approach.

Questions for authors

Q1) What methodological contribution distinguishes FLARE from standard conditional flow matching with an established rotational parameterization?

Q2) Does FLARE outperform a matched flow-matching model that predicts Cartesian displacements relative to the anchor, followed by normalization?

Q3) Beyond acceleration through surrogates, do authors envision any broader scientific capability gained through the proposed method?

4 comments
AuthorsOpenReview ↗

Author Response For Reviewer XZx7 (Part 1/2)

Author Response (Part 1/2)

We thank the reviewer for the summary of our work and for recognizing the application as interesting and previously unexplored in this setting. We also appreciate the direct questions regarding methodological novelty, generalizability, and the broader scientific value of the proposed formulation. These comments helped us sharpen both the scope of our claims and the interpretation of the new ablation results.

W1. The method combines standard conditional flow matching with established axis angle coordinates. The methodological novelty beyond adapting these components to micromagnetics is limited.

Our Response to W1. We agree that the novelty is not in a new neural architecture, a new conditional flow-matching objective, or a new axis–angle parameterization. Our contribution is instead the formulation of stochastic finite-time micromagnetic evolution as conditional transport over physically meaningful transition coordinates.

FLARE combines three design choices needed to make this formulation operational: anchor-relative finite-time rotations as the prediction target, explicit conditioning on the requested physical horizon and controls, and geometry-preserving decoding of the generated transition back to the magnetization field. The revised ablation matrix–including absolute Cartesian CFM, Cartesian residual CFM, fixed 2D tangent coordinates, Riemannian FM, and Direct U-Net—separates the effects of relative transition modeling, rotational coordinates, and the generative objective. We therefore position the contribution as a representation and problem-formulation advance for stochastic constrained-field prediction.

W2. The experiments demonstrate effectiveness within one physical system, but do not establish a generalizable methodological insight or a substantial scientific advance enabled by the approach.

Our Response to W2. We agree that the present experiments are restricted to one micromagnetic material system and do not establish cross-material or cross-domain generalization. We have made this limitation explicit in the revised Discussion.

The broader methodological relevance comes from the structure of the prediction problem itself. Beyond computational magnetism, this setting exemplifies a general generative-modeling challenge in which the state is a high-dimensional spatial field with locally manifold-constrained variables, while stochastic dynamics induce a distribution of future outcomes under identical physical controls. Micromagnetics provides a demanding instance because both geometric constraint and the conditional outcome distribution must be respected simultaneously.

Q1. What methodological contribution distinguishes FLARE from standard conditional flow matching with an established rotational parameterization?

Our Response to Q1. The distinction is in the prediction object and the role of the physical horizon. FLARE does not apply conditional flow matching directly to the future magnetization field. Given the current state msm_s, controls cc, and requested horizon Δt\Delta t, it models the distribution of the finite local rotations that carry msm_s to the future state, pθ(ω∣ms,c,Δt)p_\theta(\omega \mid m_s,c,\Delta t), and decodes the sampled rotation field through Rodrigues formula.

This makes the current state the reference for the transition coordinates, while cc and Δt\Delta t specify which finite-time transition distribution is generated. For protocols with changing controls, generated endpoints are then used as the anchors for subsequent physical segments. Thus, FLARE turns standard conditional flow matching into a finite-horizon stochastic transition model for a manifold-constrained field, instead of using it as a generator of absolute future states.

AuthorsOpenReview ↗

Author Response For Reviewer XZx7 (Part 2/2)

Author Response (Part 2/2)

Q2. Does FLARE outperform a matched flow-matching model that predicts Cartesian displacements relative to the anchor, followed by normalization?

Our Response to Q2. We thank the reviewer for suggesting this comparison. The new Cartesian residual CFM models Δm=mt−ms\Delta m = m_t - m_s and reconstructs the endpoint as m^t=normalize(ms+Δm^)\hat{m}_t = \mathrm{normalize}(m_s + \widehat{\Delta m}). The results are below (mean ± standard deviation over three seeds; added to Table 2(a)):

MethodDistribution Score ↑Angular ED ↓Fair energy score ↓Paired Angle ↓
Absolute Cartesian CFM0.7176 ± 0.007317.98 ± 0.12°0.4210 ± 0.003339.96 ± 0.29°
Cartesian residual CFM0.7480 ± 0.026717.19 ± 0.31°0.4183 ± 0.002739.16 ± 0.22°
FLARE0.7592 ± 0.005615.31 ± 0.24°0.3944 ± 0.002836.70 ± 0.23°

Across three seeds, the Cartesian residual model obtains a Distribution Score of (0.7480±0.02670.7480\pm0.0267), angular energy distance of (17.19±0.31∘17.19\pm0.31^\circ), paired angle of (39.16±0.22∘39.16\pm0.22^\circ), and fair energy score of (0.4183±0.00270.4183\pm0.0027), compared with (0.7592±0.00560.7592\pm0.0056), (15.31±0.24∘15.31\pm0.24^\circ), (36.70±0.23∘36.70\pm0.23^\circ), and (0.3944±0.00280.3944\pm0.0028) for FLARE. The residual control closes part of the gap, particularly in Distribution Score, showing that anchor-relative prediction itself is beneficial. With anchor relativity held fixed, the rotational representation provides additional improvements in the three angular/distributional metrics above. We added this control to Table 2(a) and revised the corresponding interpretation in Section 4.4.

Our Action to Q2.

On p. 9, we added a matched Cartesian residual CFM baseline to Table 2(a) and revised Section 4.4 to explain how the comparison isolates the effect of rotational coordinates while holding anchor-relative prediction fixed.

Q3. Beyond acceleration through surrogates, do authors envision any broader scientific capability gained through the proposed method?

Our Response to Q3. The primary capability beyond acceleration is efficient access to conditional outcome ensembles rather than a single predicted future. Under identical initial magnetization, physical controls, and prediction horizon, finite-temperature micromagnetic dynamics can evolve into different domain patterns and topological outcomes. In Fig. 5, we compare 30 fresh FLARE samples with 30 independent MuMax3 thermal realizations under identical inputs and observe substantial overlap in both field-level variation and physical observables such as topological charge and mean out-of-plane magnetization.

This changes the practical object of prediction from a single endpoint to a conditional distribution over physically plausible futures. Such access makes repeated stochastic interrogation substantially cheaper, enabling denser control sweeps, estimation of outcome probabilities, and sensitivity analysis over physical conditions within the validated regime. The speedup is therefore useful not only as a surrogate benchmark, but because it makes ensemble-level analysis of stochastic dynamics substantially more tractable.

We emphasize that FLARE does not introduce new physical information beyond the underlying simulator. Its scientific value is to amortize repeated stochastic simulation while preserving the conditional distributional structure needed for ensemble-level analysis.

I thank the authors for their responses.

After considering these rebuttals, and reading the reviews from other reviewers I will raise my score to weak acceptance.

I lean towards weak acceptance since rebuttals address most of the concerns however leave generalizability of the framework unclear.

AuthorsOpenReview ↗

Second Round-Author Response For Reviewer XZx7

Thank you for reconsidering our response and for thoughtful follow-up. We appreciate the remaining question regarding generalization and would like to clarify the scope of the evidence.

Within the studied micromagnetic system, the current results already provide direct evidence of generalization across held-out physical conditions and out-of-distribution settings. Cross-material and cross-domain transfer is a distinct question that we do not claim to have established in the present work, and we have made this boundary explicit in the revised Discussion.

We therefore view the present work as establishing the finite-time stochastic transition formulation and its robustness within the validated micromagnetic regime, while identifying cross-material and cross-domain validation as a natural extension. We hope this clarification resolves the ambiguity regarding the scope of generalization in the current work.

We greatly appreciate your constructive feedback, which helped us sharpen both the methodological contribution and scope of our claims.

Reviewer_VbNQOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FLARE, which represents the finite-time change of a magnetization field as anchor-relative per-cell axis-angle rotations and performs horizon-conditioned Euclidean conditional flow matching in this rotation space, so that stochastic future states at a requested time can be generated in one shot while the unit-norm constraint is preserved exactly through Rodrigues decoding. Validated on a MuMax3-generated finite-temperature dataset, FLARE outperforms seven deterministic learned baselines on distributional metrics (Distribution Score, angular energy distance, fair energy score), with ablations against Cartesian CFM, Riemannian FM, and direct regression, OOD tests on unseen current masks and temperatures, and roughly a 450–3,000× speedup over MuMax3.

Critical strengths
  1. Simple and well-motivated representation with matched internal controls. Formulating the transition as anchor-relative axis-angle rotations lets the method use standard Euclidean flow matching while guaranteeing unit-norm outputs by construction. Matched controls with the same architecture, data, and training budget (Cartesian CFM, Riemannian FM, Direct U-Net, over three seeds) show consistent gains. This matters to my assessment because it shows the improvement is not only due to backbone or training differences, although the gain over Cartesian CFM is modest.

  2. Careful and transparent evaluation. The paper uses proper scoring rules, base-condition clustered bootstrap intervals, and a simulator-only reference level that shows how far each method is from the noise floor. It also openly reports negative results, such as the mixed Stage 2 training not beating GT-only training and the higher paired angle compared with CNO-FM. This makes the reported gains more credible and is the main reason I lean towards acceptance if the baseline concerns are addressed.

Critical weaknesses

Insufficient baselines to isolate the contribution of the rotation representation. Most external comparators are deterministic, making FLARE's large improvements on distributional metrics less conclusive, although the paper also evaluates stochastic NeuralMAG and deterministic model ensembles. More importantly, the matched Cartesian CFM baseline transports absolute magnetization vectors, while FLARE predicts anchor-relative rotations. This comparison therefore conflates the effect of rotation coordinates with that of residual prediction. A Cartesian residual CFM baseline that models �0��0� and normalizes the reconstructed endpoint is essential to isolate the contribution of the rotation representation. Without this control, it remains unclear whether the reported gains arise from the proposed rotation parameterization or from the simpler benefit of anchor-relative prediction.

Questions for authors

Can the authors clarify why the matched Cartesian CFM baseline predicts absolute magnetization vectors rather than anchor-relative Cartesian residuals? Since FLARE uses anchor-relative rotations, the current comparison changes both the coordinate representation and the prediction target.

Could the authors evaluate a Cartesian residual CFM baseline that models �0��0�, reconstructs the endpoint as �1��1�, and otherwise uses the same architecture, training budget, and sampling procedure? This comparison would directly test whether FLARE's improvements come from the rotation representation or from anchor-relative prediction. The result would materially affect my assessment of the paper's central contribution.

4 comments
AuthorsOpenReview ↗

Author Response For Reviewer VbNQ (Part 1/2)

Author Response (Part 1/2)

We thank the reviewer for the careful and constructive assessment of our work. We particularly appreciate the recognition of the simple and well-motivated geometric formulation of FLARE, the matched internal controls used to separate representation and objective choices, and the careful and transparent evaluation based on proper scoring rules, base-condition-clustered uncertainty estimates, simulator-reference analyses, and explicitly reported negative results. These points closely reflect the methodological contributions and evaluation principles of the paper.

W1. Insufficient baselines to isolate the contribution of the rotation representation. Most external comparators are deterministic, making FLARE's large improvements on distributional metrics less conclusive, although the paper also evaluates stochastic NeuralMAG and deterministic model ensembles. More importantly, the matched Cartesian CFM baseline transports absolute magnetization vectors, while FLARE predicts anchor-relative rotations. This comparison therefore conflates the effect of rotation coordinates with that of residual prediction. A Cartesian residual CFM baseline that models (mt−ms)(m_t - m_s) and normalizes the reconstructed endpoint is essential to isolate the contribution of the rotation representation. Without this control, it remains unclear whether the reported gains arise from the proposed rotation parameterization or from the simpler benefit of anchor-relative prediction.

Our Response to W1. We agree that the original absolute Cartesian CFM control did not fully separate the effect of anchor-relative prediction from that of the rotational representation. We therefore added the Cartesian residual CFM control suggested by the reviewer, using the same architecture, data, training budget, and sampling procedure.

The new control sharpens the interpretation of the ablation. Moving from absolute Cartesian state prediction to an anchor-relative Cartesian residual improves all four metrics, showing that relative transition prediction is itself beneficial. Importantly, with anchor relativity held fixed, FLARE provides a further improvement in angular energy distance, fair energy score, and paired angle. The Distribution-Score difference is smaller, and the residual model shows substantially larger seed variability. We have revised the manuscript accordingly so that the contribution is attributed to both anchor-relative transition modeling and the benefit of rotational coordinates.

Our Action to W1. We revised Section 4.4 accordingly to distinguish the benefit of anchor-relative transition modeling from the additional gain provided by rotational coordinates.

Q1. Can the authors clarify why the matched Cartesian CFM baseline predicts absolute magnetization vectors rather than anchor-relative Cartesian residuals? Since FLARE uses anchor-relative rotations, the current comparison changes both the coordinate representation and the prediction target.

Our Response to Q1. The original Cartesian CFM was designed as a conventional Cartesian state-space counterpart to FLARE: it transports the target magnetization directly in Cartesian coordinates while receiving the current state msm_s as conditioning input. This comparison was intended to test rotation-space transport against absolute Cartesian state transport with the same backbone and training protocol.

We agree, however, that because FLARE represents the transition relative to msm_s, this control changes both the coordinate representation and the prediction target. The reviewer's suggested residual control isolates this distinction directly, and we have now added it to the revised manuscript.

AuthorsOpenReview ↗

Author Response For Reviewer VbNQ (Part 2/2)

Author Response (Part 2/2)

Q2. Could the authors evaluate a Cartesian residual CFM baseline that models (mt−ms)(m_t - m_s), reconstructs the endpoint as (m^t=normalize(ms+Δm^))(\hat{m}_t = \mathrm{normalize}(m_s + \widehat{\Delta m})), and otherwise uses the same architecture, training budget, and sampling procedure? This comparison would directly test whether FLARE's improvements come from the rotation representation or from anchor-relative prediction. The result would materially affect my assessment of the paper's central contribution.

Our Response to Q2. Following the reviewer's suggestion, we trained a matched Cartesian residual CFM that models Δm=mt−ms\Delta m = m_t - m_s and reconstructs the endpoint as m^t=normalize(ms+Δm^)\hat{m}_{t} = \mathrm{normalize}(m_{s} + \widehat{\Delta{m}}). All other architecture, training, and sampling choices are kept matched to FLARE. The three-seed results in Table 2(a) are:

MethodDistribution Score ↑Angular ED ↓Fair energy score ↓Paired Angle ↓
Absolute Cartesian CFM0.7176 ± 0.007317.98 ± 0.12°0.4210 ± 0.003339.96 ± 0.29°
Cartesian residual CFM0.7480 ± 0.026717.19 ± 0.31°0.4183 ± 0.002739.16 ± 0.22°
FLARE0.7592 ± 0.005615.31 ± 0.24°0.3944 ± 0.002836.70 ± 0.23°

The residual control closes part of the gap to FLARE, particularly in Distribution Score, indicating that anchor-relative prediction accounts for part of the improvement. Rotation coordinates account for most of the gain. With anchor-relative prediction held fixed, however, FLARE further reduces angular ED by 1.88°, fair energy score by 0.024 and paired angle by 2.46°. These are about 70%, 90% and 75% of FLARE's total improvement over absolute Cartesian CFM. The Distribution Score difference is smaller (0.7480 versus 0.7592), and we interpret it cautiously given the larger seed variability of the residual model.

One geometric factor that may contribute to this difference is that for small rotations, the two representations are locally similar, (Δm≈ω×ms\Delta m \approx \omega \times m_s), whereas for larger rotations the Cartesian residual norm 2sin⁡(θ/2)2\sin(\theta/2) saturates, while the rotation-vector norm remains the rotation angle (θ\theta). Rodrigues decoding also maps the predicted rotation directly back to the unit sphere. This geometric view provides a plausible explanation for the empirical advantage observed in the matched control.

Our Action to Q2. We added a matched Cartesian residual CFM baseline to Table 2(a) and revised Section 4.4 (p. 9) to distinguish the contribution of anchor-relative transition prediction from the benefit of rotational coordinates.

“The representation controls further separate the effects of anchor-relative prediction and rotational coordinates. Moving from absolute-spin Cartesian CFM to an anchor-relative Cartesian residual improves all four metrics, showing that modeling the transition relative to the current state is itself beneficial. With anchor relativity held fixed, FLARE further reduces angular energy distance from 17.19° to 15.31°, fair energy score from 0.4183 to 0.3944, and paired angular error from 39.16° to 36.70°.”


The authors have addressed most of my concerns, so I am willing to maintain my score of Weak Accept.

AuthorsOpenReview ↗

Second Round-Author Response For Reviewer VbNQ

Thank you for the follow-up and for carefully considering our response. We particularly appreciate your suggestion to include the matched Cartesian residual CFM control. This experiment helped us separate the contribution of anchor-relative prediction from the further gains of rotational coordinates and led us to revise the corresponding interpretation in the manuscript.

Together with the existing absolute-Cartesian, fixed-tangent, Riemannian-FM, and Direct-U-Net controls, the revised ablation now provides a substantially more complete representation-control comparison. We are grateful for your constructive feedback and the opportunity to strengthen the manuscript accordingly.

Reviewer_HBPTOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The papers central problem concerns stochastic micro-magnetic simulation, where thin magnetic films can evolve into many different futures under identical environmental controls. Traditional numerical solvers can only obtain these features by evolving the stochastic LLG equation using very small time increments, making them excessively expensive. Learning surrogate models on the other hand either keep the same step wise inetrgatiojn scheme or only give one deterministic outcome per condition.

FLARE reframes this problem as a conditional generation problem, where it does not directly predict the fioeld itself. Rather FALRE predicts how each local magnetization vector rotates between now and the target time. A U-Net is trained the learn the distribution of these rotation fields using conditional flow matching rather than deterministic methods, and is conditioned on physical controls.

On a new custom simulated dataset, the authors compare the prpoopsoed method with several learned baselines on multiple distributional metrics to understand the performance of FLARE, where it matches the ground truth distribution much closer than baselines.

Critical strengths

S1. The presentation is geometrically sound, The paper encodes each transition as a relative rotation which allows the flow-matching path to be exactly Euclidean, while the subsequent Rodrigues decoding is able to enforce the spherical constraint exactly. Furthermore the ablations and alternative models show that the gains exhibited by FLARE does indeed come from the rotational representations. Specifically FLARE compares against four alternative representations, namely ordinary cartesian xyz coordinates, flow matching along the surface of the sphere, tangent-plane rotation, or plain regression as opposed to flow matching. In each of these cases FLARE demonstrates improved or on par performance with greater stability in distribution metrics.

S2. The evaluation is thorough and critical of the limiations and drawbacks of the proposed method. Table 7 that FLARE results land near the mumax3 ture-on-true distribution metrics, demonstrating near floor performance when compared to the simulator. They identify a sample-size bias in their energy-distance estimator and fix the budget in Table 8. Firthermore the authors show translation-tolerance distance can sometimes hide a topology change and threrfore report the topology charge alongside. I appauld the authors for keeping the negative results prominent in the paper instead of chasing SOTA across the board.

Critical weaknesses

I think that the margins over the baselines remains relatively thin, as all the distributional results rests on only 33 held-out conditions. Additionly in Table 1 all the confidence interval of FLARE overlap with baselines methods.

Secondly the parity compared to baselines remains unclear. Notably all of the baselines except for NeuralMAG-x5 are determinsitic ones. The headline margins are taken on fair energy score that seems to penalize one-point forecast by construction (please correct me here is the authors disagree). Using the papers own 5 seed in table 15 the margin drops from 37->22% and the distribution score is statistically tied with DPOT. Additionally most of the baselines are faster than FLARE and FLARE also exhibits significant training memory and time overheads compared to baseline methods.

Questions for authors

Q1. The key selling point of the paper is one-step prediction without the integration across many smaller steps that baseline methods all leverage. However in reality one-shot generation is only practically support up to 3ns, and the main results use multiple 2ns segments with prediction 'handoffs'. One additional exp I want to see is error vs segment count and I think that the authors should use the rebuttal revision to further clarify the practical implementation limitations of the proposed method.

Q2. For baselines methods can the authors clarify what and how they are conditioned, especially the quantities and horizons are fed into the model that uses fix step integrations? Additionally did the authors retrain (and under what hyperparameters) on the new dataset? What would happen to the baselines if we match the compute that is given to FLARE as defined in the appendix?

5 comments
AuthorsOpenReview ↗

Author Response For Reviewer HBPT (Part 1/3)

We thank the reviewer for the careful reading and constructive assessment of our work. We particularly appreciate the reviewer's recognition of two aspects that are central to the paper.

First, the reviewer highlights the geometric formulation of FLARE: representing finite-time magnetization transitions as anchor-relative rotations allows conditional flow matching to operate in a Euclidean transport space, while Rodrigues decoding preserves the unit-norm constraint by construction. We also appreciate the reviewer's recognition that our matched representation and objective controls provide direct evidence for the role of the transition representation.

Second, we are grateful that the reviewer found the evaluation thorough and appropriately critical of the method's limitations. In particular, the reviewer notes our simulator-only reference level, the correction of finite-sample bias in the energy-distance analysis, the complementary topology-aware diagnostics, and our decision to retain negative and mixed results rather than selectively emphasizing favorable outcomes.

These points closely reflect the intended contributions of the paper. We address the reviewer's remaining concerns regarding statistical uncertainty, deterministic-baseline parity, temporal composition, and baseline conditioning/computational budgets below.

W1

I think that the margins over the baselines remains relatively thin, as all the distributional results rests on only 33 held-out conditions. Additionly in Table 1 all the confidence interval of FLARE overlap with baselines methods.

Our Response to W1. We agree that 33 independent base conditions limit the precision of the complete-path comparison, particularly for small margins. We would like to clarify that in Table 1 the marginal intervals do not universally overlap: Angular energy distance (AED) and fair energy score (fair ES) are separated from the strongest external baseline intervals. More importantly, because all methods are evaluated on the same physical conditions, paired base-condition comparisons are more informative than overlap between marginal intervals. In a separate single-segment evaluation over 100 held-out base conditions, FLARE's Distribution Score exceeds the strongest comparator with a paired ratio of 1.094 [1.059, 1.132] (Table 7).

Our Action to W1.

We revised the following content in the manuscript (page 8):

“FLARE has the best point estimates among the seven author-released baselines on the three distributional metrics.”

We revised the following content in the manuscript (page 24):

“The primary complete-path comparison contains 33 independent base conditions. FLARE obtains a Distribution Score of 0.762 [0.660, 0.844] and an angular energy distance of 15.30° [13.32, 17.26]. We therefore distinguish point-estimate ordering from statistical resolution: the Distribution-Score interval overlaps those of the strongest external baselines, whereas the angular-energy-distance and fair-energy intervals are separated from the strongest external comparators. Figure 4(c) additionally reports paired base-condition bootstrap ratios on the same 33 complete-path conditions.

As a complementary larger-base diagnostic, Table 7 evaluates 1,493 held-out single-segment transitions grouped into 100 independent base clusters. At the nominal bandwidth, FLARE obtains a Distribution Score of 0.8050 versus 0.7362 for the strongest comparator, corresponding to a paired FLARE/comparator ratio of 1.094 [1.059, 1.132]. The same ordering is retained across the tested bandwidths and when local translation matching is removed. These single-segment results provide higher-base-count evidence for the Distribution-Score ordering, while remaining distinct from the primary complete-path evaluation in Table 1.”

AuthorsOpenReview ↗

Author Response For Reviewer HBPT (Part 2/3)

W2

Secondly the parity compared to baselines remains unclear. Notably all of the baselines except for NeuralMAG-x5 are determinsitic ones. The headline margins are taken on fair energy score that seems to penalize one-point forecast by construction (please correct me here is the authors disagree). Using the papers own 5 seed in table 15 the margin drops from 37->22% and the distribution score is statistically tied with DPOT. Additionally most of the baselines are faster than FLARE and FLARE also exhibits significant training memory and time overheads compared to baseline methods.

Our Response to W2. We thank the reviewer for raising the important question of parity between stochastic FLARE forecasts and deterministic baselines. A deterministic prediction can be viewed as a degenerate predictive distribution and can therefore be evaluated with the same proper fair energy score. However, we agree that advantage persists once deterministic models are given multiple predictive members.

We previously reported five-model deep ensembles for four external baselines and the matched Direct U-Net control (Table 16). At the same-five-output evaluation budget, ensembling substantially narrows the gap but does not eliminate it. The strongest external ensemble, Poseidon-T, reaches a fair energy score of 0.503 and an angular energy distance 18.93°, compared with 0.393 and 15.30° for FLARE. The best ensemble Distribution Score is 0.744 for DPOT-Ti versus 0.762 for FLARE, while Poseidon-T obtains a lower paired angle (33.31° versus 36.72°). Thus, the fair-energy advantage decreases from 37.3% against the strongest single-checkpoint baseline to 21.8% against the strongest five-model ensemble, and we now make this distinction explicit in the manuscript.

These ensemble comparisons match the number of predictive outputs rather than total training compute. In fact, the five-member Poseidon-T, CNO-FM, and DPOT-Ti ensembles require approximately 1.2–2.3× the training device-hours of FLARE. We do not interpret this as an exact single-model compute-matched experiment, but as a complementary parity test showing that the distributional advantage is not solely a consequence of comparing a stochastic model with a one-point forecast.

As an additional baseline-favoring stress test, we also randomize deterministic prediction by perturbing the initial anchor and sweep the perturbation amplitude without retraining (Appendix Table 17). Even the best point estimate across this oracle sweep reaches a fair energy score of 0.502, above FLARE 0.393.

We also agree that FLARE is slower than most learned baselines and requires more training time and memory. The revised manuscript makes this tradeoff explicit: FLARE's advantage is distributional fidelity for stochastic future-state prediction, while several deterministic surrogates are substantially faster at inference. FLARE is 448× faster than MuMax3 for a single query and 3,062× faster in batched inference, while lying on the quality–speed Pareto front (Fig. 2c). Heun-5 further halves inference latency with essentially unchanged angular error (Fig. 7). Under the cost assumptions in Appendix D, the one-time training cost is recovered after approximately 123 queries relative to MuMax3.

Our Action to W2.

We revised the following content in the manuscript (page 8):

“FLARE also retains the best distributional point estimates against five-model ensembles of deterministic baselines (Table 16), although ensembling narrows its fair-energy margin to 21.8% and Poseidon-T attains a lower paired angle.”

We added the following contents in the manuscript (page 34):

“Deterministic single checkpoints define valid one-point predictive distributions and can therefore be evaluated by the same proper fair energy score. Nevertheless, to test whether FLARE's advantage is primarily due to comparing a stochastic predictor with one-point forecasts, we additionally construct five-model ensembles from independently trained deterministic checkpoints. This provides a matched five-output evaluation.”

“For context, the five-member Poseidon-T, CNO-FM, and DPOT-Ti ensembles correspond to approximately 1.22×, 2.25×, and 2.27× FLARE's Stage-1 training device-hours, respectively, based on Table 13.”

“As an additional deliberately baseline-favoring sensitivity test, we randomize deterministic predictions by perturbing the initial drive-start anchor.”

“Because the perturbation amplitude is selected retrospectively from the full sweep, we treat these minima as an oracle sensitivity bound rather than as a tuned baseline result.”

AuthorsOpenReview ↗

Author Response For Reviewer HBPT (Part 3/3)

Q1

The key selling point of the paper is one-step prediction without the integration across many smaller steps that baseline methods all leverage. However in reality one-shot generation is only practically support up to 3 ns, and the main results use multiple 2 ns segments with prediction “handoffs”. One additional exp I want to see is error vs segment count and I think that the authors should use the rebuttal revision to further clarify the practical implementation limitations of the proposed method.

Our Response to Q1. We thank the reviewer for the valuable suggestion to examine prediction quality as the number of segments increases.

We first clarify the temporal protocol. The primary complete-path experiment contains one predicted-state handoff: a driven segment is followed by a single one-shot 3.5-ns relaxation prediction. The fixed “2 ns + 3 ns” protocol is used only as the standardized timing benchmark. In the saved-frame horizon audit, legal exact single-segment targets are available through 3.0 ns; the 3.5-ns relaxation is nevertheless evaluated directly in the primary complete-path task. In principle, FLARE can also generate predictions at longer time horizons.

Following the reviewer’s suggestion, we added a fixed-total-duration segmentation experiment (Fig. 15). The driven segment is held fixed, while the relaxation is divided into 1–9 equal intervals, yielding 2–10 FLARE calls in total. Across this range, Distribution Score remains within [0.750, 0.761], AED within [15.05°, 15.86°], and fair ES within [0.391, 0.405]. We therefore observe no systematic net degradation as the number of segments increases under this segmentation test.

Our Action to Q1.

We added Figure 15 (p. 37) to evaluate how prediction quality changes as the number of relaxation segments increases while the driven segment is held fixed.

Q2

For baselines methods can the authors clarify what and how they are conditioned, especially the quantities and horizons are fed into the model that uses fix step integrations? Additionally did the authors retrain (and under what hyperparameters) on the new dataset? What would happen to the baselines if we match the compute that is given to FLARE as defined in the appendix?

Our Response to Q2. We thank the reviewer for asking us to clarify baseline conditioning, retraining, and computational parity. We have expanded the comparator protocol in Appendix Table 6 and Appendix A.4.

All methods receive the same underlying physical controls, while retaining their native conditioning and temporal interfaces. FLARE uses scalar and spatial controls together with an explicit physical horizon. Poseidon-T and CNO-FM receive normalized physical-control maps and query the requested endpoint through their native time-conditioning mechanism. DPOT-Ti, MPP-AViT-Ti, and PDEArena advance autoregressively in 0.25-ns steps, with the total duration represented by the step sequence and fractional tail. LE-PDE advances in 0.25-ns latent steps and re-encodes at control switches, while NeuralMAG incorporates the controls in the LLG solver and handles the requested duration through RK4 integration. Appendix Table 6 summarizes these interfaces explicitly.

All learned-dynamics comparators are adapted to the same x5 training split for 50,000 updates and evaluated at the final checkpoint. Poseidon-T, CNO-FM, DPOT-Ti, and MPP-AViT-Ti are initialized from author-released pretrained weights, whereas PDEArena and LE-PDE are trained from scratch. Appendix Table 6 reports the optimizer, learning rate, temporal interface, update count, and measured training device-hours; Appendix A.4 further documents the common bring-up and checkpoint-selection protocol.

The primary comparison therefore matches adaptation updates, while training device-hours vary across architectures. As a complementary higher-compute comparison, the five-member Poseidon-T, CNO-FM, and DPOT-Ti ensembles use approximately 1.22×, 2.25×, and 2.27× FLARE's Stage-1 training device-hours, respectively. FLARE still achieves lower angular energy distance and fair energy score in these comparisons (Appendix Table 13 and 16). This provides additional evidence that the distributional advantage is not explained simply by a larger training budget.

Our Action to Q2.

We revised the following content in the manuscript (page 7):

“Each method receives the same underlying physical controls through its native conditioning and temporal interface (Appendix Table 6).”

We added Table 6 in Appendix A.4 (p. 21) to document the inputs, conditioning, temporal interface, and training settings of each method.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.