acceptodds
Under review as a conference paper at ICLR 2027

From Fixed Grids to Moving Particles: A Transferable Latent Operator for Fluid Dynamics

Abstract

Lagrangian modeling is vital to fluid dynamics, as it characterizes particle transport and complements the Eulerian description. However, Lagrangian trajectories are less commonly available than Eulerian fields, while most neural operators are trained and evaluated primarily in the Eulerian representation. This mismatch motivates a new learning problem: can a model trained solely on Eulerian observations generalize zero-shot from Eulerian field prediction to Lagrangian particle rollout, without Lagrangian supervision or task-specific adaptation? To address this problem, we propose the Transferable Latent Operator (TLO), which learns a unified flow representation shared by Eulerian field prediction and Lagrangian particle rollout. TLO decouples latent flow evolution from coordinate-dependent decoding: querying the evolving latent representation at fixed spatial coordinates yields Eulerian fields, whereas querying velocities at particle positions and recursively updating these positions enables Lagrangian rollout. Across five fluid-dynamics benchmarks, TLO achieves strong Eulerian forecasting accuracy and competitive zero-shot Lagrangian rollout, with further gains from decoder fine-tuning using Lagrangian supervision.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
AI for science, Physical Intelligence, Fluid dynamics, Latent representation

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

does a model trained only on eulerian fields support closed loop particle rollout? TLO combines a local global encoder, a latent processor, and a coordinate conditioned decoder to answer this questions. The same model forecasts fields on a fixed grid and supplies velocities at moving particle positions. The evaluation covers five field prediction and three particle rollout benchmarks, with additional experiments on decoder adaptation and prediction beyond the supervised grid support.

Critical strengths
  • The Eulerian results are strong. TLO achieves the lowest reported error on all five benchmarks in Table 2, and the ablations support combining local and global information. This is useful evidence that the proposed representation works well for field forecasting.

  • I found the distinction between Eul, Ref, and Path informative. The results show that better grid predictions do not necessarily produce better particle trajectories, which is an important point for evaluating these models. I also appreciate that the authors keep Eulerian-only training and decoder adaptation seperate.

Critical weaknesses
  1. My main hesitation is the contribution beyond existing coordinate-query operators. UPT and LNO already separate latent processing from decoding at requested spatial locations. TLO adds a local-global patch representation and a Transolver-inspired processor, but Table 3 only evaluates these choices through Eulerian error. It remains unclear which components actually help with moving-particle inference. The explicit transfer evaluation is useful, and I do not assume that prior work tested exactly this protocol. Still, I would find the contribution more convincing if the architectural choices were connected more directly to improvements in Ref and Path.

  2. The particle results leave this question open. In Table 4, Transolver has lower Ref and Path errors on Burgers3D and SEA, despite TLO’s better Eulerian scores. Direct decoding and high-resolution bilinear interpolation also perform very similarly in Table 12. The beyond-grid comparision could provide a stronger distinction but interpolation is simply undefined after grid exit, and other coordinate decoders are not evaluated there. This shows that TLO can continue producing predictions, but does not establish a particular accuracy advantage. The authors acknowledge these limitations, which I appreciate. Given the overlap with existing methods, however, I am not yet convinced that the demonstrated benefit for particle rollout is substantial enough for acceptance.

Questions for authors
  1. Additional evidence: which architectural components are responsible for the particle-transfer performance? Could the local/global and processor ablations also report Ref and Path? Particle results for GINO, which is already included in the Eulerian comparison, would also help assess the contribution.

  2. Clarification: does Table 12 use an Eulerian-only or decoder-adapted checkpoint? Are direct-decoding versus interpolation results available from the same Eulerian-only checkpoint across all three particle benchmarks, using identical integration and boundary handling?

  3. Additional evidence: how does TLO compare beyond the grid support with another coordinate-query operator, such as UPT or LNO, or with interpolation equipped with an explicit extrapolation rule? A comparison under the same Eulerian-only supervision, reporting outside-region errors separately, could materially change my assessment.

2 comments

Thank you for the careful review and helpful suggestions. We appreciate your recognition of TLO’s Eulerian forecasting performance and the value of distinguishing Eul, Ref, and Path.

  1. Contribution beyond existing coordinate-query operators

We agree that TLO shares query-independent latent processing and coordinate-conditioned decoding with UPT/LNO. Our contribution, however, is to formulate zero-shot Eulerian-to-Lagrangian transfer, unifying field prediction and particle simulation, and introduce Eul/Ref/Path to measure cross-representation transfer. This combines a controlled, systematic study of EO transfer with an architecture designed for particle rollout.

To address transfer task, TLO couples a Local Perceiver with global aggregation through domain decomposition and uses attention and FFN to capture latent flow dynamic. Our architectural contribution lies in this particular combination of local–global representation and latent processing.

As you pointed out, the Eul-only ablations in Table 3 do not establish which components help particle prediction. We have therefore added Ref/Path results and extended the ablations to DAM, NS, and TGV. The results show that the Local Perceiver is essential for accurate particle rollout in TLO: compared with the variant without Local component, the full model reduces Path error by 35.1%, 67.7%, and 93.5%, respectively. Removing Global or FFN also increases Ref and Path. These results directly connect the architectural choices to particle-transfer performance. In the tables, Improvement denotes the error reduction of the full model relative to each ablated variant: : $\text{Improvement}(%)= (E_{\text{Baseline}}-E_{\text{Full}} )/E_{\text{Baseline}}\times100%$.

DatasetComponentsEulRefPathImprovement Eul (%)Improvement Ref (%)Improvement Path (%)
DAM2DFull3.12E-032.41E-033.72E-03———
w/o Global3.70E-034.12E-035.62E-0315.6841.5033.81
w/o Local4.19E-033.97E-035.73E-0325.5439.2935.08
w/o FFN3.75E-033.65E-035.09E-0316.8033.9726.92
NS2DFull1.98E-054.33E-054.14E-04———
w/o Global2.22E-055.29E-054.83E-0410.81%18.15%14.29%
w/o Local4.21E-056.09E-051.28E-0352.97%28.90%67.66%
w/o FFN4.51E-055.48E-057.34E-0456.10%20.99%43.60%
TGV3DFull1.75E-044.78E-047.39E-03———
w/o Global1.90E-044.98E-047.53E-037.89%4.02%1.86%
w/o Local7.08E-037.82E-031.13E-0197.53%93.89%93.46%
w/o FFN1.96E-045.14E-047.83E-0310.71%7.00%5.62%
  1. Particle-transfer baselines

We have added GINO’s particle-prediction results and completed the NS2D and native TGV3D evaluations, extending particle evaluation to all five datasets. The expanded comparison shows TLO achieves strong Ref and Path results on these flow datasets.

EO modelDAMNSTGVBurgersSEA
EulRefPathEulRefPathEulRefPathEulRefPathEulRefPath
UPT2.38E-027.62E-021.44E-011.31E-031.31E-032.49E-027.54E-038.34E-031.26E-015.59E-032.91E-037.73E-038.97E-014.92E-021.06E-02
LNO3.75E-034.35E-035.30E-038.73E-051.57E-045.73E-037.54E-038.33E-031.26E-014.42E-033.36E-039.42E-035.21E-015.29E-021.10E-02
GINO6.04E-035.62E-037.56E-033.68E-058.32E-054.42E-032.47E-045.83E-048.15E-035.35E-034.83E-035.38E-035.90E-015.11E-021.23E-02
DeepLag1.65E-024.08E-024.65E-024.04E-033.56E-033.86E-023.17E-045.92E-048.82E-033.38E-032.16E-037.39E-038.17E-015.01E-021.08E-02
Transolver6.13E-037.70E-039.60E-031.40E-041.45E-042.70E-037.67E-041.11E-031.62E-024.22E-043.36E-042.75E-035.22E-012.34E-028.52E-03
TLO3.12E-032.41E-033.72E-031.98E-054.33E-054.14E-041.75E-044.78E-047.39E-039.88E-056.17E-043.01E-031.77E-013.00E-028.99E-03
  1. EO readout comparisons

The original Table 12 used a decoder-adapted checkpoint. We now provide EO Direct comparisons on DAM2D, Burgers3D, and SEA. For each dataset, Direct and Interp use the same EO checkpoint and Eulerian feedback, with identical integration and boundary handling as specified in Table 1. Only the particle-velocity readout changes. In the resolution experiments, we vary only the sampling density of the interpolation grid. These results show how readout resolution affects particle prediction under the same model and particle rollout settings.

DatasetResolutionInterp RefDirect RefInterp PathDirect Path
DAM2D32×642.34E-032.41E-033.98E-033.72E-03
DAM2D16×323.29E-032.41E-036.05E-033.72E-03
DAM2D8×161.13E-022.41E-032.49E-023.72E-03
DAM2D4×83.56E-022.41E-036.13E-023.72E-03
Burgers3D32×32×326.19E-056.17E-042.02E-033.01E-03
Burgers3D16×16×161.93E-046.17E-042.45E-033.01E-03
Burgers3D8×8×81.16E-036.17E-043.06E-033.01E-03
Burgers3D4×4×43.21E-036.17E-044.96E-033.01E-03
SEA180×3002.08E-023.00E-028.70E-038.99E-03
SEA90×1502.13E-023.00E-028.75E-038.99E-03
SEA60×1002.22E-023.00E-028.81E-038.99E-03
SEA45×752.34E-023.00E-028.87E-038.99E-03
SEA30×502.67E-023.00E-029.00E-038.99E-03
SEA18×303.57E-023.00E-029.32E-038.99E-03
  1. Beyond-grid comparisons

We have also added EO beyond-grid comparisons with coordinate-query operators and grid-based baselines equipped with explicit extrapolation.

We have strengthened Section 6.3 with both coordinate-query and grid-based baselines. We retrained LNO, GINO, and Transolver using the same in-domain data and training settings as TLO. LNO, GINO and TLO use direct coordinate queries to predict velocities outside the grid.

Transolver uses bilinear interpolation inside the grid. Outside, constant extrapolation extends the boundary values, while linear extrapolation uses the slope between the two outermost grid layers. These baselines allow a direct comparison of beyond-grid accuracy.

  1. TLO achieves the lowest outside-region Path error under EO supervision. Compared with LNO and GINO, TLO reduces Outside@20 Path error by 85.7% and 83.2%, respectively. Transolver’s errors vary substantially with the extrapolation rule: constant extrapolation gives it the lowest outside-region Ref, but its Path remains higher than TLO’s.
  2. The extrapolation rule substantially affects Transolver’s errors, and the distinction between Ref and Path. TLO achieves more accurate closed-loop particle prediction, demonstrating its advantage in Eulerian-to-Lagrangian transfer.
  3. Decoder adaptation (DA) provides the largest improvements for TLO, reducing outside-region Ref and Path errors by 71.7% and 42.0%, respectively. TLO-DA achieves the lowest errors among the adapted models and outperforms both Transolver extrapolation baselines, further demonstrating the adaptability of TLO's learned representation for beyond-grid particle prediction.

Together with the component ablations, these results demonstrate that TLO’s local–global design supports effective Eulerian-to-Lagrangian transfer, achieving accurate closed-loop particle prediction even beyond the Eulerian output grid.

We hope these additional results clarify our contributions and address the reviewer's concerns.

Training SettingModelRead-outIn-grid@10 Ref ↓In-grid@10 Path ↓Outside@20 Ref ↓Outside@20 Path ↓Full@20 Path ↓
EOLNODirect1.02E-023.78E-021.56E-014.23E-018.05E-02
EOGINODirect1.32E-021.51E-021.64E-013.60E-019.59E-02
EOTransolverLinear8.82E-031.57E-022.89E-012.50E-014.94E-02
EOTransolverConstant8.82E-031.57E-029.12E-022.37E-014.88E-02
EOTLODirect3.22E-041.89E-041.35E-016.05E-024.48E-03
DALNODirect1.02E-023.78E-021.48E-014.23E-018.04E-02
DAGINODirect1.32E-021.51E-021.14E-013.59E-019.59E-02
DATLODirect3.22E-041.89E-043.82E-023.51E-022.51E-03
Reviewer_f8RAOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes the Transferable Latent Operator, which aims to enable zero-shot transfer from Eulerian field prediction to Lagrangian particle rollout without requiring particle-level supervision or task-specific adaptation. The proposed framework consists of a local-global patch encoder, an attention-based latent processor, and a coordinate-conditioned decoder. By decoupling latent flow evolution from coordinate-dependent decoding, TLO supports both fixed-grid Eulerian forecasting and moving-coordinate particle queries within a unified framework. Experiments across five fluid dynamics benchmarks demonstrate strong Eulerian forecasting performance and competitive zero-shot Lagrangian rollout. The authors further investigate decoder adaptation and particle prediction beyond the fixed Eulerian grid support.

Critical strengths
  1. Strong Eulerian forecasting performance.The method is evaluated on five fluid dynamics benchmarks covering two-dimensional and three-dimensional flows, numerical simulations, and ocean reanalysis data. As shown in Table 2, TLO achieves the lowest Eulerian forecasting errors among the compared methods across all five benchmarks, demonstrating its effectiveness in modeling flow dynamics.
  2. Systematic evaluation of cross-representation prediction. The paper evaluates several neural operators under a unified Eulerian-only training setting and introduces separate metrics for Eulerian field prediction, velocity prediction at reference particle positions, and closed-loop particle trajectory errors. This evaluation provides useful insights into the relationship between field prediction accuracy and particle rollout performance.
Critical weaknesses

Although the paper evaluates Lagrangian rollout on several benchmarks, TGV3D contains native three-dimensional particle trajectories but is only evaluated for Eulerian forecasting. Previous studies, such as LagrangeBench, have demonstrated the feasibility of Lagrangian rollout evaluation on this dataset. While Burgers3D already provides a three-dimensional Lagrangian evaluation, its reference trajectories are generated through numerical integration of Eulerian velocity fields. Additional evaluation on TGV3D would provide complementary evidence of the method's effectiveness on native three-dimensional particle trajectories.

Questions for authors

Could the authors clarify why Lagrangian Ref and Path evaluations are not reported on TGV3D, given that the dataset contains native particle trajectories and has been used for Lagrangian rollout evaluation in previous studies? Such results would further strengthen the evaluation of zero-shot Eulerian-to-Lagrangian transfer in three-dimensional fluid dynamics.

1 comment

Thank you for recognizing TLO’s Eulerian forecasting performance and our systematic transfer evaluation. We also appreciate your suggestion to include native 3D particle trajectories for a more complete evaluation.

To our knowledge, this is the first study to introduce a joint Eul/Ref/Path evaluation of cross-representation prediction. We therefore sought to evaluate this task across diverse settings. As summarized in Table 1, we selected datasets based on their native representation (Lagrangian SPH or Eulerian grids), dimensionality, data source, and physical properties (free surfaces, compressibility, or dissipation).

These considerations also guided the original Ref/Path evaluation. In 2D, we chose DAM2D for its native SPH trajectories and evolving free surface, an important feature of particle simulation. In 3D, Burgers3D provided a complementary setting with reference trajectories derived from Eulerian fields. We also included SEA to extend the evaluation beyond simulations to ocean reanalysis data. Our aim was to cover a broader range of data sources and flow conditions in the Eul/Ref/Path evaluation.

However, as you rightly point out, this selection left out native three-dimensional particle trajectories. These were available in TGV3D but omitted from our original Ref/Path evaluation. . We have now added TGV3D and NS2D Ref/Path results, extending particle evaluation to all five datasets. We have also included GINO as a particle-prediction baseline. These additions make the zero-shot Eulerian-to-Lagrangian transfer evaluation more complete.

EO modelDAMNSTGVBurgersSEA
EulRefPathEulRefPathEulRefPathEulRefPathEulRefPath
UPT2.38E-027.62E-021.44E-011.31E-031.31E-032.49E-027.54E-038.34E-031.26E-015.59E-032.91E-037.73E-038.97E-014.92E-021.06E-02
LNO3.75E-034.35E-035.30E-038.73E-051.57E-045.73E-037.54E-038.33E-031.26E-014.42E-033.36E-039.42E-035.21E-015.29E-021.10E-02
GINO6.04E-035.62E-037.56E-033.68E-058.32E-054.42E-032.47E-045.83E-048.15E-035.35E-034.83E-035.38E-035.90E-015.11E-021.23E-02
DeepLag1.65E-024.08E-024.65E-024.04E-033.56E-033.86E-023.17E-045.92E-048.82E-033.38E-032.16E-037.39E-038.17E-015.01E-021.08E-02
Transolver6.13E-037.70E-039.60E-031.40E-041.45E-042.70E-037.67E-041.11E-031.62E-024.22E-043.36E-042.75E-035.22E-012.34E-028.52E-03
TLO3.12E-032.41E-033.72E-031.98E-054.33E-054.14E-041.75E-044.78E-047.39E-039.88E-056.17E-043.01E-031.77E-013.00E-028.99E-03
Reviewer_SQcPOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether a model trained exclusively on Eulerian fields sampled on fixed grids can directly support Lagrangian particle trajectory prediction. The authors propose the Transferable Latent Operator (TLO), which combines a local–global encoder, an adaptive slice-based latent processor, and a coordinate-conditioned decoder to predict flow fields. During inference, the model autoregressively predicts Eulerian fields at fixed coordinates while querying velocities at recursively updated particle positions and numerically integrating them to obtain trajectories. Experiments cover five Eulerian forecasting datasets, three particle trajectory datasets, and rollout beyond the spatial support of the Eulerian output grid.

Critical strengths
  1. Separating fixed-grid field error (Eul), velocity error at reference trajectory positions (Ref), and closed-loop trajectory error (Path) helps distinguish errors arising from field prediction, spatial querying, and particle-position feedback. The observation that lower Eul does not necessarily imply lower Path is also empirically informative.

  2. In Table 2, TLO achieves the lowest Eul across all five datasets, covering two- and three-dimensional systems, simulated flows, and ocean reanalysis data. The matched-query-budget ablations in Table 3 also support the value of combining local and global information.

  3. The authors distinguish the zero-shot Eulerian-only (EO) setting from decoder adaptation (DA) using particle-velocity supervision. Appendix B explicitly acknowledges that being evaluable outside the grid does not imply accurate extrapolation, and that direct decoding is not guaranteed to outperform interpolation within the grid. These qualifications help readers interpret the results appropriately.

Critical weaknesses
  1. The emphasized design of “output-query-independent latent processing plus coordinate-conditioned decoding” is closely related to existing approaches. UPT already adopts an encode–latent evolution–decode framework with decoding at arbitrary spatial locations, while LNO explicitly supports decoding beyond the training sampling locations, including interpolation and extrapolation.

  2. In the EO results in Table 4, TLO performs well on DAM2D, but Transolver achieves lower Ref and Path errors on both Burgers3D and SEA. The results therefore provide more direct evidence that TLO can perform zero-shot particle rollout and achieves accurate Eulerian forecasting. The evidence that its architecture improves trajectory transfer remains limited.

  3. Section 6.3 and Table 6 compare TLO against standard interpolation that becomes undefined after particles leave the grid. This establishes a difference in interface availability, but is insufficient to demonstrate TLO’s predictive advantage outside the grid. In particular, coordinate-query baselines such as UPT and LNO are already included elsewhere in the paper but are absent from the beyond-grid comparison. Including at least one of these methods, or equipping interpolation with an explicit, simple extrapolation rule, would be more informative than only reporting that interpolation cannot continue.

  4. The model still predicts the Eulerian grid at every step and feeds the predicted field back into the encoder to advance the flow state. Consequently, the computational cost of particle rollout must include both the Eulerian feedback loop and the additional particle queries.

Questions for authors

See weakness.

2 comments

Thank you for the careful review. We appreciate the positive assessment of our Eul/Ref/Path metrics, local–global ablations, and EO–DA comparison. Your comments help us clarify the architectural contribution and add the evidence for particle prediction.

  1. Contribution beyond existing coordinate-query operators

We agree that TLO shares query-independent latent processing and coordinate-conditioned decoding with UPT/LNO. Our contribution, however, is to formulate zero-shot Eulerian-to-Lagrangian transfer, unifying field prediction and particle simulation, and introduce Eul/Ref/Path to measure cross-representation transfer. This combines a controlled, systematic study of EO transfer with an architecture designed for particle rollout.

To address transfer task, TLO couples a Local Perceiver with global aggregation through domain decomposition and uses attention and FFN to capture latent dynamic. Our architectural contribution lies in this particular combination of local–global representation and latent processing.

We have extended the Eul/Ref/Path ablations to DAM, NS, and TGV. The results show that the Local component is essential for accurate particle rollout in TLO: compared with the variant without Local component, the full model reduces Path error by 35.1%, 67.7%, and 93.5%, respectively. Removing Global or FFN also increases Ref and Path, directly linking these architectural choices to cross-representation prediction accuracy. In the tables, Improvement denotes the error reduction of the full model relative to each ablated variant: : $\text{Improvement}(%)= (E_{\text{Baseline}}-E_{\text{Full}} )/E_{\text{Baseline}}\times100%$.

DatasetComponentsEulRefPathImprovement Eul (%)Improvement Ref (%)Improvement Path (%)
DAM2DFull3.12E-032.41E-033.72E-03———
w/o Global3.70E-034.12E-035.62E-0315.68%41.50%33.81%
w/o Local4.19E-033.97E-035.73E-0325.54%39.29%35.08%
w/o FFN3.75E-033.65E-035.09E-0316.80%33.97%26.92%
NS2DFull1.98E-054.33E-054.14E-04———
w/o Global2.22E-055.29E-054.83E-0410.81%18.15%14.29%
w/o Local4.21E-056.09E-051.28E-0352.97%28.90%67.66%
w/o FFN4.51E-055.48E-057.34E-0456.10%20.99%43.60%
TGV3DFull1.75E-044.78E-047.39E-03———
w/o Global1.90E-044.98E-047.53E-037.89%4.02%1.86%
w/o Local7.08E-037.82E-031.13E-0197.53%93.89%93.46%
w/o FFN1.96E-045.14E-047.83E-0310.71%7.00%5.62%
  1. Expanded particle evaluation and the effect of interpolation

We agree that strong Eulerian results alone do not establish particle-prediction gains. We have therefore added particle evaluations on NS2D and native TGV3D trajectories, bringing the evaluation to all five datasets. TLO achieves strong Ref and Path results on these two flows.

EO modelNSTGV
EulRefPathEulRefPath
UPT1.31E-031.31E-032.49E-027.54E-038.34E-031.26E-01
LNO8.73E-051.57E-045.73E-037.54E-038.33E-031.26E-01
GINO3.68E-058.32E-054.42E-032.47E-045.83E-048.15E-03
DeepLag4.04E-033.56E-033.86E-023.17E-045.92E-048.82E-03
Transolver1.40E-041.45E-042.70E-037.67E-041.11E-031.62E-02
TLO1.98E-054.33E-054.14E-041.75E-044.78E-047.39E-03

The lower Ref/Path errors you noted for Transolver led us to examine its velocity readout. Alongside the original (tri- or bi-) linear results in Table 4, we now report nearest-neighbor interpolation on Burgers3D and SEA. Switching to nearest-neighbor increases both errors, showing that interpolation affects particle metrics even when the field-prediction model is unchanged.

Burgers3DSEA
ModelInterpolationRefPathRefPath
TransolverLinear3.36E-042.75E-032.34E-028.52E-03
TransolverNearest6.53E-045.37E-032.38E-021.51E-02
TLODirect6.17E-043.01E-033.00E-028.99E-03
  1. Beyond-grid prediction

We have added coordinate-query and grid-based baselines to Section 6.3. LNO, GINO, and Transolver were retrained on the same in-domain data and with the same training settings as TLO. TLO, LNO, and GINO use direct coordinate queries outside the grid. Transolver uses bilinear interpolation inside the grid. Outside, constant extrapolation extends the boundary values, while linear extrapolation uses the slope between the two outermost grid layers. This allows us to compare beyond-grid accuracy against methods that also support prediction outside the grid.

  1. Under EO supervision, TLO achieves the lowest outside-region Path error, with reductions of 85.7% relative to LNO and 83.2% relative to GINO. Constant extrapolation gives Transolver the lowest outside-region Ref, but its Path remains higher than TLO’s.
  2. The extrapolation rule substantially affects Transolver’s errors, and the distinction between Ref and Path。
  3. DA improves all three coordinate-query models, with the largest gains for TLO: outside-region Ref and Path decrease by 71.7% and 42.0%, respectively. TLO-DA achieves the lowest errors among the adapted models and also outperforms both Transolver extrapolation baselines across all reported metrics.
SettingModelRead-outIn-grid@10 Ref ↓In-grid@10 Path ↓Outside@20 Ref ↓Outside@20 Path ↓Full@20 Path ↓
EOLNODirect1.02E-023.78E-021.56E-014.23E-018.05E-02
EOGINODirect1.32E-021.51E-021.64E-013.60E-019.59E-02
EOTransolverLinear8.82E-031.57E-022.89E-012.50E-014.94E-02
EOTransolverConstant8.82E-031.57E-029.12E-022.37E-014.88E-02
EOTLODirect3.22E-041.89E-041.35E-016.05E-024.48E-03
DALNODirect1.02E-023.78E-021.48E-014.23E-018.04E-02
DAGINODirect1.32E-021.51E-021.14E-013.59E-019.59E-02
DATLODirect3.22E-041.89E-043.82E-023.51E-022.51E-03

Together with the component ablations, TLO’s specific design helps it learn flow representations from Eulerian fields that remain useful for particle rollout beyond the grid.

  1. Computational cost

Thank you for pointing out the full cost of particle rollout. You are right that the Eulerian feedback loop at every step, including re-encoding, processing, and querying, must be counted.

We retain this loop for fair comparison: the coordinate-query models (TLO, UPT, LNO, and GINO) and grid-based models (Transolver and DeepLag) follow the same autoregressive field-prediction protocol, taking an Eulerian grid as input at each step. The feedback cost follows from this shared training protocol.

An alternative is to train the processor for multi-step latent-space rollout, avoiding repeated full-grid decoding and re-encoding while retaining particle-velocity queries. This would require a corresponding training objective and rollout scheme. We appreciate this point because it brings an important connection into focus: efficient decoding must be supported by a training protocol matched to the intended inference scheme.

Comments on the paper

1 comment

We thank the Area Chair and reviewers for their careful assessment and constructive feedback. We also thank the organizers for providing PAT through the ICLR submission system. Its feedback helped us check our evaluation scripts and improve the manuscript.

We would first like to reiterate the central motivation and contribution of our work. Our paper investigates a fundamental question: Can a neural operator trained exclusively on Eulerian flow fields at fixed spatial coordinates generalize to closed-loop Lagrangian particle prediction without any particle-level supervision? To the best of our knowledge, this is the first study to explicitly formulate and systematically investigate zero-shot Eulerian-to-Lagrangian transfer under Eulerian-only supervision.

Motivated by this question, we propose TLO as a unified framework, which couples a Local Perceiver with global aggregation through domain decomposition and uses physical attention and FFN blocks to capture latent flow information. We further introduce three complementary metrics—Eul, Ref, and Path—to systematically evaluate the model’s transfer ability. Our contribution therefore goes beyond arbitrary-coordinate querying itself: we aim to establish a systematically study Eulerian-to-Lagrangian transfer under a unified training and evaluation framework.

Across the reviews, TLO’s strong Eulerian forecasting results and the value of distinguishing Eul, Ref, and Path error are consistently recognized. The central remaining question is whether TLO’s specific architectural choices provide demonstrable benefits for particle prediction beyond the ability to query arbitrary coordinates. We address this question through two targeted experiments:

  1. Architectural ablations. We extend the local–global–FFN ablations to Ref and Path, directly investigating which components contribute to particle-transfer performance. The Local Perceiver plays a critical role in TLO’s particle-rollout accuracy: relative to the variant without Local, the full model reduces Path error by 35.1% on DAM, 67.7% on NS, and 93.5% on TGV. Removing Global or FFN also increases Ref and Path. These results show that the architectural benefits extend beyond Eulerian forecasting to particle prediction.
  2. Beyond-grid predictive accuracy. We compare TLO against coordinate-query models and grid-based models equipped with explicit extrapolation under the same Eulerian-only (EO) supervision, evaluating predictive accuracy beyond the grid. TLO achieves the lowest outside-region Path error among the evaluated EO configurations, outperforming LNO, GINO, and Transolver with linear extrapolation. TLO’s value therefore lies in learning flow representations that transfer from Eulerian supervision to closed-loop particle prediction beyond the grid, not simply in providing an arbitrary-coordinate query interface.

Following the reviewers’ suggestions, we have also added GINO’s particle results and extended particle evaluation to all five datasets. The new EO Direct–Interp comparisons use the same checkpoint, integration scheme, and boundary handling for each dataset. The readout experiments show that interpolation method and readout-grid resolution affect Ref and Path, clarifying the role of readout choice in particle evaluation. Decoder adaptation (DA) brings further improvements, which we report separately from EO transfer. We also clarify that the rollout cost includes the Eulerian feedback loop.

Together, these additions strengthen the paper’s central contributions: the formulation of zero-shot Eulerian-to-Lagrangian transfer, a local–global operator that improves particle-rollout accuracy, and a joint Eul/Ref/Path evaluation covering both in-grid and beyond-grid prediction. The new experiments show why TLO’s architecture matters for particle prediction and support its value beyond arbitrary-coordinate queryability.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.