acceptodds
Under review as a conference paper at ICLR 2027

RAL-MoE-ROM: Hierarchical Mixture-of-Experts for Regime-Adaptive Reduced-Order Model

Abstract

Reduced-order models (ROMs) approximate high-dimensional dynamical systems using a small number of state variables. However, their predictive accuracy can deteriorate when parameter changes lead to different dynamical regimes. Different regimes may require different reduced representations and dynamics, making it difficult for a single model to remain accurate across the full parameter range. We introduce RAL-MoE-ROM, a hierarchical mixture-of-experts framework with two levels: regime-local specialists and cross-regime routing with fusion. At the first level, each local model, called a specialist, uses a structured sparse mixture of experts to improve its reduced dynamics and is trained through multi-step autonomous rollouts. At the second level, a parameter-only outer router provides preferences within prescribed adjacent specialist pairs, supporting Top-1 selection or history-conditioned specialists fusion. When two specialists are selected, a learned gate uses the initial physical history to combine their reconstructed fields. Across three flow benchmarks spanning steady, Hopf-onset, and periodic regimes, the local specialists provide accurate autonomous predictions, while cross-regime routing further improves prediction in the evaluated regime overlaps.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Reduced-order model, Mixture-of-experts, Regime-adaptive localized model, Cross-regime Routing

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_veApOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

RAL-MoE-ROM is a novel reduced order approximation method that combines regime-local POD–Galerkin models with sparse mixture-of-experts corrections and history-conditioned, physical-space fusion. The authors provide experiments that cover three fluid flow geometries; cross-regime fusion is tested on prescribed steady–Hopf and Hopf–periodic overlaps.

Critical strengths
  1. Interesting approach to an important problem. I find it clever that each specialist uses its own POD basis. This way, fusion after physical reconstruction avoids aligning incompatible local reduced coordinates.
  2. The author provide a broad comparison accros a range of methods: POD–Galerkin, OpInf, recurrent, FNO, ql-ROM. Also, architecture ablations provide useful evidence for the reported gains (Tables 1–2 and 6).
Critical weaknesses
  1. The biggest concern in my opinion is that authors perform manual regime pairing (i.e., overlapping regions and candidate specialist pairs are prescribed), so fully automatic regime selection is not demonstrated (Section 4.2). Ideally, the system should determine which specialist or combination of specialists is applicable to a new physical parameter. This significatly weakens the methods.
  2. Despite Table 3, I find the validation narrow. For instance, each interface uses only two Reynolds numbers and overlapping windows. A much extensive range of prameter should be considered to provide a clearer picture of the method.
  3. The inference cost remains quite high. It looks like the periodic specialist takes 3.32x the Dense baseline's inference time despite sparse activation (Table 12). This is particularly concerning given that reduced-order modelling is motivated by computational efficiency.
Questions for authors

These are related to the weaknesses above, in the same order:

  1. Can the full system identify applicable specialists without manually supplied pairs?
  2. Do the gains in Table 3 persist across more unseen Reynolds numbers and longer rollouts?
  3. Is Table 12 a typo ? If not, then did the authors consider improving the accuracy and computational efficiency compared to denser baselines?
6 comments

We sincerely thank the reviewer for identifying three central questions concerning automatic specialist selection, the breadth of validation, and computational efficiency. The revised experiments provide evidence that RAL-MoE-ROM can determine specialist participation without manually supplied interface-specific pairs, retain aggregate predictive benefits over broader parameter evaluations and longer rollouts, and achieve lower measured inference latency than the matched Dense control.

To address manual regime pairing, we removed the interface-specific pair mask and evaluated E2 selection of one or two eligible specialists from the constructed regime-local model family. With identical frozen specialists and T2-C gates, E2 achieves the lowest mean joint error in all 15 benchmark–horizon comparisons against Reynolds-number thresholding and validation-based selection. This demonstrates that automatic specialist selection contributes to prediction accuracy, rather than merely assigning weights to a prescribed pair.

The evaluation now covers 38 Circular, 38 Square, and 44 Pinball Reynolds-number configurations, with autonomous routing comparisons extending to K = 512 on Square. E2 retains the lowest mean joint error throughout these comparisons. A separate study evaluates fusion on twelve Square interface trajectories at K = 512, with four extended to K = 1024. At K = 512, T2-C improves on both individual specialists at H–P, while Hopf remains more accurate at S–H; the expanded results therefore support broader aggregate predictive benefits while identifying the interface dependence of long-horizon fusion.

Regarding computational efficiency, the original 3.32× ratio resulted from inconsistent timing scopes rather than a typographical error, and we have removed that comparison. With the revised implementation and matched outer prediction configurations on Square, RAL-MoE-ROM is faster than the total-parameter-matched Dense control, including a 22.71% latency reduction under adaptive T2-C at K = 48. The separate accuracy comparison gives approximately 16.0% and 32.4% lower velocity and pressure errors at the same horizon, supporting a favorable accuracy–cost trade-off. We provide the supporting experiments and manuscript locations in the detailed responses below, and for more detailed responses, please refer to the supplementary material Reply to Reviewer veAp.pdf.

Comment 1: The biggest concern in my opinion is that authors perform manual regime pairing (i.e., overlapping regions and candidate specialist pairs are prescribed), so fully automatic regime selection is not demonstrated (Section 4.2). Ideally, the system should determine which specialist or combination of specialists is applicable to a new physical parameter. This significatly weakens the methods.

Question: Can the full system identify applicable specialists without manually supplied pairs?

Response: Thank you for highlighting this concern. Within the available model family, the revised evaluation selects applicable specialists without manually supplied interface-specific pairs. We now evaluate E2 without the interface-specific pair mask: it determines which eligible specialists participate and whether prediction uses one or two. The regime-local POD representations and chart-assigned training labels remain physically prescribed; the added experiment establishes automatic selection within this model family, rather than automatic discovery of regime boundaries.

For μ=Re\mu=\mathrm{Re}, E2 provides preferences πi(μ)\pi_i(\mu). Among eligible specialists, it selects the highest-ranked one when the maximum preference is at least τ=0.90\tau=0.90, and otherwise selects the two highest-ranked specialists (Section 3.4, Eq. (18), p. 6). Eligibility requires the projected operators and initialization from the prescribed physical history. A single specialist predicts directly. When an adjacent pair is selected, both specialists encode the same initial history in their own coordinates and evolve independently; their corresponding frozen T2-C gate combines the reconstructed fields. Selection is parameter-only, while the fusion weight also uses the initial three physical states. The selected set and weight remain fixed throughout the rollout (Section 3.4 and Appendix C).

We compare E2 with two alternatives while retaining identical frozen specialists and pair-specific T2-C gates. Reynolds-number thresholding uses fixed transition centers a,ba,b and half-widths w,vw,v: it selects S--H in [a−w,a+w][a-w,a+w], H--P in [b−v,b+v][b-v,b+v], and a single specialist elsewhere. Candidate rules first maximize the number of fully evaluable parameters and then minimize the equally weighted mean joint error across Reynolds numbers. Validation-based selection instead measures each admissible configuration's mean autonomous prediction error at calibration parameters, selects the minimum-error configuration there, and assigns it to a target parameter by nearest-neighbor lookup. Calibration and evaluation trajectories are separate. Appendix F.6, Eqs. (66)--(70) and Table 11 (p. 26), define both baselines. The gates retain their original E2 pair-preference and initial-history inputs, so this comparison changes specialist participation, not the fusion mechanism.

E2 selects two specialists for 27/38 Circular, 27/38 Square, and 30/44 Pinball configurations, corresponding to 71.1%71.1\%, 71.1%71.1\%, and 68.2%68.2\% (Table 12, p. 27). More importantly, it has the lowest mean full-window joint error, Ejoint=Eu+EpE_{\mathrm{joint}}=E_u+E_p, in all 15 benchmark--horizon comparisons (Table 13, p. 28). These results demonstrate predictive value beyond simply activating two specialists: E2 coordinates independent local ROMs more effectively than the two tested selection alternatives, without requiring a manually supplied interface-specific pair.

Comment 2: Despite Table 3, I find the validation narrow. For instance, each interface uses only two Reynolds numbers and overlapping windows. A much extensive range of prameter should be considered to provide a clearer picture of the method.

Question: Do the gains in Table 3 persist across more unseen Reynolds numbers and longer rollouts?

Response: Thank you for highlighting the narrow original validation. We now distinguish the controlled fixed-pair fusion comparison from a broader automatic-selection evaluation and a separate long-horizon interface study. This allows us to assess broader predictive gains without treating additional overlapping windows as independent trajectories.

The original 16 windows per interface comprise eight overlapping windows from each of two Reynolds-number trajectories. Holding specialist predictions fixed isolates the fusion rule: T2-C reduces joint error from 1.5786%1.5786\% to 1.0954%1.0954\% at S--H and from 1.6424%1.6424\% to 0.8123%0.8123\% at H--P, approximately 30.6%30.6\% and 50.5%50.5\% reductions relative to E2 Top-1 (Table 3, p. 8). It also outperforms equal blending, E2-probability weighting, and training-fitted constant-weight and logit-offset controls. The architecture-matched history ablation gives 39.64%39.64\% and 16.85%16.85\% reductions relative to the history-ablated gate, supporting the value of initial-history information (Appendix F.1, Figure 2, pp. 18--19).

The expanded evaluation covers 38 Circular, 38 Square, and 44 Pinball parameter configurations. The autonomous horizons are K∈{24,48,56,128}K\in\{24,48,56,128\}, K∈{24,48,56,128,256,512}K\in\{24,48,56,128,256,512\}, and K∈{20,32,56,96,128}K\in\{20,32,56,96,128\}, respectively. Each geometry uses its own specialists and gates; within each benchmark, the three selection methods share identical frozen specialists and T2-C gates. E2 has the lowest mean joint error in all 15 benchmark--horizon comparisons. Its longest-horizon errors are 6.3343%6.3343\%, 6.7456%6.7456\%, and 0.5712%0.5712\%, respectively (Section 4.2; Appendix F.6, Table 13, p. 28). Equal output-step counts do not imply equal physical durations across benchmarks.

The separate direct fusion study uses twelve Square trajectories, six per interface, at K=512K=512 (T=2048T=2048); four extend to K=1024K=1024 (T=4096T=4096). At K=512K=512, T2-C's H--P joint error is 8.68%8.68\%, below Periodic (8.92%8.92\%) and Hopf (16.24%16.24\%). At S--H, the errors are 4.83%4.83\% for T2-C, 5.42%5.42\% for Steady, and 3.05%3.05\% for Hopf. Thus, H--P fusion gains persist in this study, while Hopf remains preferable at S--H. Appendix F.5, Table 10, and Figures 6--7 (pp. 23--25) report the direct comparisons and trajectory-resolved energy and nonlinear-residual diagnostics.

The expanded evidence supports history-conditioned fusion beyond simple weighting and the aggregate predictive contribution of E2 over broader parameter evaluations and longer autonomous rollouts. We report the interface dependence of direct fusion separately, rather than equating the routing comparison with uniform superiority over the best individual specialist. This provides a clearer account of where coordinated local ROMs improve prediction.

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

This paper proposes RAL-MoE-ROM, a hierarchical mixture-of-experts framework for reduced-order modeling across different flow regimes. Each regime-specific POD–Galerkin specialist is augmented with structured sparse MoE corrections and trained through autonomous rollouts. An outer parameter-conditioned router and a history-conditioned fusion gate coordinate neighboring specialists in physical space, avoiding the need for a shared reduced representation. Experiments on three CFD benchmarks demonstrate improved local prediction accuracy, while cross-regime fusion further reduces errors in the evaluated overlap regions.

Critical strengths
  1. The framework addresses both representation mismatch and dynamical differences across regimes. In particular, fusing predictions in physical space provides a principled way to combine independently constructed ROMs.

  2. The proposed specialists outperform several competitive baselines, including Dense, Structured, and Global alternatives. The circular-cylinder ablations provide convincing evidence for the effectiveness of the local sparse correction mechanism.

  3. Comparisons against fixed-weight, parameter-only, and history-ablated fusion help establish the value of history conditioning. The additional analyses of physical residuals, energy, and computational cost are also appreciated.

Critical weaknesses
  1. The regime partitions and adjacent specialist pairs are predefined, while the outer router is trained using chart-assigned labels. Consequently, the experiments primarily demonstrate improved fusion within prescribed pairs, rather than autonomous regime identification or specialist selection. This weakens the central claim of regime adaptivity.

  2. The main quantitative fusion experiments are confined to the centered-square benchmark, involving only two held-out Reynolds numbers and 16 rollout windows per interface. These windows may be correlated, while the reported near-zero variation across gate-training seeds does not establish generalization across independent trajectories. The effectiveness of fusion over longer horizons and other geometries remains unclear.

  3. Local MoE and outer fusion are evaluated largely through separate experiments, without a unified ablation establishing their combined necessity. The explicit polynomial branches are not independently ablated, and the Hopf specialist consistently selects fixed expert pairs. Moreover, sparse inference is 3.32× slower than Dense, raising concerns about whether simpler localized ensembles could achieve comparable performance.

Questions for authors
  1. Can the authors evaluate routing without predefined specialist pairs, and compare E2 against simple Reynolds-number thresholding or validation-based selection? This would help clarify the actual contribution of learned regime routing.

  2. Can the fusion experiments be extended to more unseen Reynolds numbers, independent trajectories, and other geometries? How does history-conditioned fusion behave over substantially longer rollouts, particularly regarding nonlinear physical residuals?

  3. Can the authors provide a unified ablation comparing Dense, Structured, and sparse-MoE specialists with and without outer fusion, including removal of the polynomial branches? Reporting accuracy alongside inference cost would better establish the necessity of the complete hierarchy.

6 comments

We sincerely thank the reviewer for the constructive comments on regime adaptivity, fusion evaluation, and the contribution of the two-level architecture. The additional experiments provide complementary evidence for automatic specialist selection, the value of initial-history information, and the predictive and computational benefits of the complete framework. We summarize the main findings below before addressing each comment in detail.

We evaluated E2 without an interface-specific pair mask, allowing it to determine both which specialists participate and whether one or two are selected. With the specialists and T2-C gates held fixed, E2 achieves the lowest mean joint error among the three selection methods in all 15 benchmark–horizon comparisons. This supports the predictive contribution of automatic selection within the constructed regime-local model family, rather than only improved fusion within prescribed pairs.

To broaden the evaluation beyond the original overlapping windows, we added parameter evaluations covering 38 Circular, 38 Square, and 44 Pinball configurations, with Square routing comparisons extending to K = 512. The architecture-matched history ablation shows that retaining initial-history information reduces the full-window joint error by 39.64% at S–H and 16.85% at H–P relative to the history-ablated gate. We distinguish these controlled measurements of history conditioning from the additional trajectory evaluations, rather than treating gate-training seed variation as evidence of trajectory-level generalization.

We further added trajectory-resolved physical diagnostics up to K = 1024. The analysis explicitly separates fusion-induced energy and nonlinear residual defects from total residuals and prediction errors, revealing decreasing S–H defects after the early transient and larger late-time H–P defects. These results characterize the interface-dependent physical effects of fusion alongside the expanded prediction comparisons.

The architectural ablation shows that, with outer routing and T2-C enabled, RAL-MoE-ROM achieves lower velocity and pressure errors than Dense, Structured, and the variant without polynomial correction branches at all three evaluated horizons. The expanded Hopf analysis also demonstrates actual within-group expert switching: with the model frozen and its sole group fixed, the 13 completed test trajectories at K = 512 exhibit 38 velocity-pair switches and 33 pressure-pair switches. Thus, the fixed pairs observed in the original short windows are not a structural restriction of the specialist.

Finally, after correcting the original timing-scope mismatch, matched inference measurements favor RAL-MoE-ROM over the total-parameter-matched Dense control, including a 22.71% latency reduction under adaptive T2-C at K = 48. Together, the accuracy ablations, routing diagnostics, and timing measurements support the complementary contributions of internal sparse routing, explicit correction branches, and history-conditioned fusion. We present the detailed evidence and corresponding manuscript revisions in the responses below, and for more detailed responses, please refer to the supplementary material Reply to Reviewer Huko.pdf.

Comment 1: The regime partitions and adjacent specialist pairs are predefined, while the outer router is trained using chart-assigned labels. Consequently, the experiments primarily demonstrate improved fusion within prescribed pairs, rather than autonomous regime identification or specialist selection. This weakens the central claim of regime adaptivity.

Question: Can the authors evaluate routing without predefined specialist pairs, and compare E2 against simple Reynolds-number thresholding or validation-based selection? This would help clarify the actual contribution of learned regime routing.

Response: Thank you for highlighting this distinction. We now evaluate E2 without the interface-specific pair mask: it determines which eligible specialists participate and whether prediction uses one or two. The regime-local POD representations and chart-assigned training labels remain physically prescribed; the added experiment establishes automatic selection within this model family, rather than automatic discovery of regime boundaries.

For μ=Re\mu=\mathrm{Re}, E2 provides preferences πi(μ)\pi_i(\mu). Among eligible specialists, it selects the highest-ranked one when the maximum preference is at least τ=0.90\tau=0.90, and otherwise selects the two highest-ranked specialists (Section 3.4, Eq. (18), p. 6). Eligibility requires the projected operators and initialization from the prescribed physical history. A single specialist predicts directly. When an adjacent pair is selected, both specialists encode the same initial history in their own coordinates and evolve independently; their corresponding frozen T2-C gate combines the reconstructed fields. Selection is parameter-only, while the fusion weight also uses the initial three physical states. The selected set and weight remain fixed throughout the rollout (Section 3.4 and Appendix C).

We compare E2 with two alternatives while retaining identical frozen specialists and pair-specific T2-C gates. Reynolds-number thresholding uses fixed transition centers a,ba,b and half-widths w,vw,v: it selects S--H in [a−w,a+w][a-w,a+w], H--P in [b−v,b+v][b-v,b+v], and a single specialist elsewhere. Candidate rules first maximize the number of fully evaluable parameters and then minimize the equally weighted mean joint error across Reynolds numbers. Validation-based selection instead measures each admissible configuration's mean autonomous prediction error at calibration parameters, selects the minimum-error configuration there, and assigns it to a target parameter by nearest-neighbor lookup. Calibration and evaluation trajectories are separate. Appendix F.6, Eqs. (66)--(70) and Table 11 (p. 26), define both baselines. The gates retain their original E2 pair-preference and initial-history inputs, so this comparison changes specialist participation, not the fusion mechanism.

E2 selects two specialists for 27/38 Circular, 27/38 Square, and 30/44 Pinball configurations, corresponding to 71.1%71.1\%, 71.1%71.1\%, and 68.2%68.2\% (Table 12, p. 27). More importantly, it has the lowest mean full-window joint error, Ejoint=Eu+EpE_{\mathrm{joint}}=E_u+E_p, in all 15 benchmark--horizon comparisons (Table 13, p. 28). These results demonstrate predictive value beyond simply activating two specialists: E2 coordinates independent local ROMs more effectively than the two tested selection alternatives, without requiring a manually supplied interface-specific pair.

Comment 2: The main quantitative fusion experiments are confined to the centered-square benchmark, involving only two held-out Reynolds numbers and 16 rollout windows per interface. These windows may be correlated, while the reported near-zero variation across gate-training seeds does not establish generalization across independent trajectories. The effectiveness of fusion over longer horizons and other geometries remains unclear.

Question: Can the fusion experiments be extended to more unseen Reynolds numbers, independent trajectories, and other geometries? How does history-conditioned fusion behave over substantially longer rollouts, particularly regarding nonlinear physical residuals?

Response: Thank you for this important distinction. We expanded both parameter-level evaluation and long-horizon physical diagnostics. The original 16 windows per interface come from two Reynolds-number trajectories, so gate-training seed variation measures optimization variability on fixed windows, not generalization across independent trajectories.

The architecture-matched history ablation reduces mean full-window joint error by 39.64%39.64\% at S--H and 16.85%16.85\% at H--P relative to the history-ablated gate (Appendix F.1 and Figure 2, pp. 18--19). This supports the contribution of initial-history information beyond the parameter and E2 preference. Separately, the expanded evaluation covers 38 Circular, 38 Square, and 44 Pinball configurations using benchmark-specific specialists and gates. E2 has the lowest mean joint error among the three selection methods in all 15 benchmark--horizon comparisons; the longest-horizon errors are 6.3343%6.3343\% at K=128K=128 for Circular, 6.7456%6.7456\% at K=512K=512 for Square, and 0.5712%0.5712\% at K=128K=128 for Pinball (Appendix F.6, Table 13, p. 28). Equal KK does not imply equal physical duration across benchmarks.

The separate Square fusion study evaluates twelve trajectories, six per interface, at K=512K=512 (T=2048T=2048); four extend to K=1024K=1024 (T=4096T=4096), 21.33 and 42.67 times the original T=96T=96 interval. The fusion weight is determined from the initial history and remains fixed. Common offline operators measure interior-wake residuals, while field errors and kinetic energy use the common physical mesh (Appendices F.3--F.5, pp. 22--25).

We subtract the weighted specialist residual fields from the fused residual before taking the RMS, rather than subtracting residual norms. With β=1−α\beta=1-\alpha and δu=u1−u2\delta u=u_1-u_2, the additional defects satisfy ΔR^u=−αβND(δu,δu)\Delta\widehat R_u=-\alpha\beta N_D(\delta u,\delta u) and ΔR^p=−αβdiv⁡DND(δu,δu)\Delta\widehat R_p=-\alpha\beta\operatorname{div}_D N_D(\delta u,\delta u), where NDN_D is the diagnostic bilinear convective operator (Eqs. (63)--(64)). The normalized energy deficit measures the difference between weighted specialist energies and fused energy relative to reference energy; it is distinct from the fused prediction's energy error.

S--H defects decrease after the early transient, whereas H--P defects grow and fluctuate later. Mean normalized energy deficits at K=512K=512 are 0.01%0.01\% and 0.04%0.04\% at S--H and H--P; at K=1024K=1024, the late-quarter H--P mean is 0.07%0.07\% (Table 9, p. 24; Figures 6--7, p. 25). Equation (65) separates the additional defect's squared norm from its interaction with inherited residuals, explaining why nonzero defects can increase or decrease the total residual.

At K=512K=512, T2-C's H--P joint error is 8.68%8.68\%, versus 8.92%8.92\% for Periodic and 16.24%16.24\% for Hopf. At S--H, Hopf is more accurate: 3.05%3.05\% versus 4.83%4.83\% for T2-C. Hopf also has the lowest pressure residual among predictions at both interfaces (Table 10, p. 25). We therefore distinguish field accuracy, total residuals, and fusion-induced defects.

Together, the history ablation and expanded routing comparison support predictive benefits beyond the original overlap windows. The long-horizon diagnostics make the physical effects of fusion explicit and trajectory-resolved, while preserving independent specialist evolution; they characterize finite-horizon, interface-dependent behavior rather than guaranteeing asymptotic stability.

Reviewer_fbQaOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper builds separate reduced-order models for different flow regimes (steady, Hopf onset, periodic). Each model adds a sparse MoE correction to a POD–Galerkin model and is trained with multi-step rollouts. A small router on Re picks one model, or blends two neighboring models in physical space using a gate based on the initial flow history. Tests on three 2D flows show lower error than POD–Galerkin and other baselines.

Critical strengths
  1. Clear way to combine local models. Each regime model keeps its own POD space, and predictions are blended only in physical space. This avoids mapping between different reduced coordinates, which is a real problem for local ROMs.

  2. Careful ablations. The paper compares against dense and structured controls with matched active parameters, a global ROM, and several fusion rules, including a history-ablated gate. These show that both the MoE correction and the history-based gate help in the tested cases.

Critical weaknesses
  1. The regimes are given by hand. The regime regions, overlaps, and candidate pairs are all fixed in advance, and the router only sees Re. The hard part of being “regime-adaptive” is not learned.

  2. Easy benchmarks and short horizons. All tests are 2D, low-Re, laminar flows with rollouts of only 24–56 steps. There is no long-time stability test, which is a key concern for learned ROMs.

  3. Fusion results rest on very few cases. T2-C is tested on only 16 windows from two Re values per overlap, and it is almost the same as the convex oracle. It is unclear if the gain holds more broadly.

  4. Weak baselines and high cost. OpInf (114% error) and ql-ROM look poorly tuned, and baselines get less training than the specialists. The model is also heavy (31–48M parameters for 32 modes) and 3.3× slower than the dense control.

Questions for authors
  1. How does T2-C perform on more parameter values and longer rollouts, for example several hundred steps?

  2. Can you report OpInf and ql-ROM with the same tuning and training budget as your specialists? Their very high errors suggest weak implementations.

6 comments

We sincerely thank the reviewer for the careful assessment and constructive suggestions. The revised manuscript provides stronger evidence for automatic specialist selection, accurate prediction over extended autonomous rollouts, and a favorable accuracy–cost trade-off. We summarize the principal revisions and findings before presenting our detailed responses.

To address the concern about manual regime pairing, we removed the interface-specific pair mask and evaluated E2 selection of one or two eligible specialists within the constructed regime-local model family. With identical frozen specialists and T2-C gates, E2 achieves the lowest mean joint error among the three selection methods in all 15 benchmark–horizon comparisons. These results demonstrate the predictive contribution of automatic specialist selection beyond Reynolds-number thresholding and validation-based selection.

We expanded the parameter evaluation to 38 Circular, 38 Square, and 44 Pinball Reynolds-number configurations. The routing comparison extends to K = 512 on Square, where E2 achieves a mean joint error of 6.7456%, while a separate trajectory-resolved fusion study extends to K = 1024. These experiments broaden the evidence beyond the original short-window evaluations and show that the aggregate advantage of E2 selection persists over the tested prediction horizons.

The controlled fusion experiments further support the contribution of initial-history information. T2-C reduces the full-window joint error relative to E2 Top-1 by approximately 30.6% at S–H and 50.5% at H–P, while outperforming the simpler weighting controls. Its terminal-step errors also approach those of the a posteriori convex oracle despite determining the fusion weight using only information available at initialization.

We also expanded the OpInf regularization search and documented 96 validation evaluations for the ql-ROM adaptation. At K = 48 on Circular Periodic, the proposed specialist achieves a joint error of 2.3998%, compared with 28.6920% for ql-ROM and 114.9897% for OpInf. Both baselines undergo validation-based model selection, and the revised discussion makes their implementation scope and differences in tuning budgets explicit rather than claiming an equal-budget comparison.

Finally, we corrected the timing-scope mismatch underlying the original 3.32× comparison and evaluated matched outer prediction configurations on Square. With the revised implementation, RAL-MoE-ROM has lower measured latency than the total-parameter-matched Dense control, including a 22.71% reduction under adaptive T2-C at K = 48. The separate architectural accuracy comparison also favors RAL-MoE-ROM at all three evaluated horizons, supporting practical benefits in both prediction accuracy and inference cost. We provide the supporting protocols, results, and manuscript locations in the point-by-point responses below, and for more detailed responses, please refer to the supplementary material Reply to Reviewer fbQa.pdf.

Comment 1: The regimes are given by hand. The regime regions, overlaps, and candidate pairs are all fixed in advance, and the router only sees Re. The hard part of being ``regime-adaptive'' is not learned.

Response: Thank you for highlighting this distinction. We now evaluate E2 without the interface-specific pair mask: it determines which eligible specialists participate and whether prediction uses one or two. The regime-local POD representations and chart-assigned training labels remain physically prescribed; the added experiment establishes automatic selection within this model family, rather than automatic discovery of regime boundaries.

For μ=Re\mu=\mathrm{Re}, E2 provides preferences πi(μ)\pi_i(\mu). Among eligible specialists, it selects the highest-ranked one when the maximum preference is at least τ=0.90\tau=0.90, and otherwise selects the two highest-ranked specialists (Section 3.4, Eq. (18), p. 6). Eligibility requires the projected operators and initialization from the prescribed physical history. A single specialist predicts directly. When an adjacent pair is selected, both specialists encode the same initial history in their own coordinates and evolve independently; their corresponding frozen T2-C gate combines the reconstructed fields. Selection is parameter-only, while the fusion weight also uses the initial three physical states. The selected set and weight remain fixed throughout the rollout (Section 3.4 and Appendix C).

We compare E2 with two alternatives while retaining identical frozen specialists and pair-specific T2-C gates. Reynolds-number thresholding uses fixed transition centers a,ba,b and half-widths w,vw,v: it selects S--H in [a−w,a+w][a-w,a+w], H--P in [b−v,b+v][b-v,b+v], and a single specialist elsewhere. Candidate rules first maximize the number of fully evaluable parameters and then minimize the equally weighted mean joint error across Reynolds numbers. Validation-based selection instead measures each admissible configuration's mean autonomous prediction error at calibration parameters, selects the minimum-error configuration there, and assigns it to a target parameter by nearest-neighbor lookup. Calibration and evaluation trajectories are separate. Appendix F.6, Eqs. (66)--(70) and Table 11 (p. 26), define both baselines. The gates retain their original E2 pair-preference and initial-history inputs, so this comparison changes specialist participation, not the fusion mechanism.

E2 selects two specialists for 27/38 Circular, 27/38 Square, and 30/44 Pinball configurations, corresponding to 71.1%71.1\%, 71.1%71.1\%, and 68.2%68.2\% (Table 12, p. 27). More importantly, it has the lowest mean full-window joint error, Ejoint=Eu+EpE_{\mathrm{joint}}=E_u+E_p, in all 15 benchmark--horizon comparisons (Table 13, p. 28). These results demonstrate predictive value beyond simply activating two specialists: E2 coordinates independent local ROMs more effectively than the two tested selection alternatives, without requiring a manually supplied interface-specific pair.

Comment 2: Easy benchmarks and short horizons. All tests are 2D, low-Re, laminar flows with rollouts of only 24--56 steps. There is no long-time stability test, which is a key concern for learned ROMs.

Question: How does T2-C perform on more parameter values and longer rollouts, for example several hundred steps?

Response: Thank you for emphasizing accumulated error over longer autonomous predictions. We expanded the parameter coverage and rollout horizons within the three two-dimensional flow benchmarks, and E2 retains the lowest mean joint error among the three selection methods throughout the extended comparison.

The evaluation covers 38 Circular, 38 Square, and 44 Pinball Reynolds-number configurations. Horizons are K∈{24,48,56,128}K\in\{24,48,56,128\} for Circular, K∈{24,48,56,128,256,512}K\in\{24,48,56,128,256,512\} for Square, and K∈{20,32,56,96,128}K\in\{20,32,56,96,128\} for Pinball. Each benchmark uses its own specialists and gates. E2, Reynolds-number thresholding, and validation-based selection share identical frozen specialists and T2-C gates, isolating the effect of specialist participation.

E2 has the lowest mean full-window joint error in all 15 benchmark--horizon comparisons. At the longest horizon in this routing study, the errors are 6.3343%6.3343\% for Circular (K=128K=128), 6.7456%6.7456\% for Square (K=512K=512), and 0.5712%0.5712\% for Pinball (K=128K=128). For Square, the corresponding K=512K=512 errors are 6.9046%6.9046\% for thresholding and 6.7927%6.7927\% for validation-based selection. The results are summarized in Section 4.2 (p. 9) and reported in Appendix F.6 and Table 13 (p. 28). Here KK counts autonomous output steps, not integration stages; benchmark-dependent snapshot intervals mean that equal KK does not imply equal physical duration (Appendix D).

A separate interface-focused study evaluates frozen T2-C on twelve Square trajectories at K=512K=512 (T=2048T=2048), with four extended to K=1024K=1024 (T=4096T=4096). These durations are 21.33 and 42.67 times the original T=96T=96 interval. At K=512K=512, H--P fusion has joint error 8.68%8.68\%, compared with 8.92%8.92\% for Periodic and 16.24%16.24\% for Hopf. At S--H, Hopf remains more accurate: 3.05%3.05\%, compared with 4.83%4.83\% for T2-C and 5.42%5.42\% for Steady. Appendix F.5, Tables 9--10, and Figures 6--7 (pp. 23--25) report field errors, energy effects, and nonlinear residuals separately.

Together, these experiments extend the evidence beyond short overlapping windows: automatic specialist selection retains its aggregate prediction advantage over the tested horizons, while the longer interface study characterizes where fusion complements individual specialists. The results support finite-horizon predictive robustness within the evaluated flows, rather than a formal guarantee of asymptotic stability.

Comments on the paper

1 comment

We sincerely thank the Area Chair and all three reviewers for their careful assessment and constructive feedback. The revised manuscript strengthens the evidence for the central contribution of RAL-MoE-ROM: coordinating independently constructed regime-local hybrid ROMs through automatic specialist selection and history-conditioned physical-space fusion. The additional results support broader autonomous prediction accuracy, input-dependent internal expert selection, and a favorable measured accuracy–cost trade-off.

The shared concern about manual regime pairing is addressed by evaluating E2 without an interface-specific pair mask. Within the physically constructed local model family, E2 determines which eligible specialists participate and whether prediction uses one specialist or two. With identical frozen specialists and T2-C gates, it achieves the lowest mean joint error against Reynolds-number thresholding and validation-based selection in all 15 benchmark–horizon comparisons. These results support the predictive contribution of automatic specialist participation beyond fusion within prescribed pairs, as reported in Section 4.2 and Appendix F.6.

The expanded evaluation covers 38 Circular, 38 Square, and 44 Pinball Reynolds-number configurations, with routing comparisons extending to K = 512 on Square. Separate trajectory-resolved fusion diagnostics extend to K = 1024 and quantify energy changes, nonlinear residual defects, and field errors. The architecture-matched history ablation additionally shows 39.64% and 16.85% reductions in full-window joint error relative to the history-ablated gate at S–H and H–P. These complementary experiments broaden the predictive evidence and identify the interface dependence of long-horizon fusion, rather than assuming that fusion uniformly improves every diagnostic or always outperforms the best individual specialist.

The contribution of the specialist architecture is examined with outer routing and T2-C enabled for every variant. RAL-MoE-ROM achieves lower velocity and pressure errors than Dense, Structured, and the variant without polynomial correction branches at K = 24, 48, and 56. Expanded Hopf diagnostics further demonstrate expert-pair switching with frozen model parameters and a fixed group: the 13 completed test trajectories at K = 512 exhibit 38 velocity-pair switches and 33 pressure-pair switches. These results support both the predictive contribution of the architecture and its input-dependent internal routing behavior, as detailed in Appendices F.7 and G.1.

We also clarified the baseline-selection procedures by expanding the OpInf regularization search and documenting 96 validation evaluations for the ql-ROM adaptation. The proposed specialist retains substantially lower prediction errors in the reported Circular comparison. Both baselines are validation-selected, with their implementation scope and differences in tuning budgets stated explicitly; the comparison does not claim matched training budgets.

For computational efficiency, we corrected the timing-scope mismatch underlying the original 3.32× ratio and evaluated matched outer prediction configurations on Square. With the revised implementation, RAL-MoE-ROM has lower measured latency than the total-parameter-matched Dense control, including a 22.71% reduction under adaptive T2-C at K = 48. The separate architectural accuracy comparison also gives approximately 16.0% and 32.4% lower velocity and pressure errors at this horizon. These results, reported in Section 4.4 and Appendices F.7 and H, demonstrate measured inference savings alongside improved accuracy.

Taken together, the revisions support a framework that preserves local reduced representations and projected dynamics while coordinating their predictions automatically. We believe the expanded experiments strengthen the evidence for the predictive and computational value of RAL-MoE-ROM, and we have provided detailed point-by-point responses and corresponding manuscript revisions for all three reviewers.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.