acceptodds
Under review as a conference paper at ICLR 2027

DAFL: Disturbance-aware Field Learning for Robust CFD Warm-Starting

Abstract

Aerodynamic design optimization necessitates repeated analyses across diverse geometries and flow conditions, each requiring costly computational fluid dynamics (CFD) simulations. Learning-based warm-starting offers a promising way to reduce solver iterations by predicting initial states closer to convergence. However, existing methods can fail on complex three-dimensional aerodynamic flows, where spatially heterogeneous disturbances and highly imbalanced flow variable distributions lead to inaccurate or even invalid predictions that undermine solver convergence. We propose Disturbance-aware Field Learning (DAFL), a framework for robust CFD warm-starting. DAFL adaptively partitions the flow field according to disturbance characteristics and performs disturbance-aware sampling and regression, enabling accurate modeling of critical disturbances. Geometry and flow conditions are injected throughout the backbone to capture case-dependent variations, while logarithmic wall distance improves the representation of steep near-wall velocity gradients. For flow variables with highly imbalanced distributions, DAFL performs regression in log space and maps predictions back through an exponential transformation, improving optimization while preserving physically valid initial states. Experiments on complex three-dimensional aircraft data demonstrate that DAFL consistently outperforms existing approaches in flow field prediction. More importantly, DAFL enables robust CFD warm-starting, substantial solver acceleration, and strong generalization across diverse geometries and flow conditions.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Disturbance-aware Field Learning, Flow Field Prediction, Learning-based Warm-starting, Computational Fluid Dynamics

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_YD1EOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DAFL, a learning based framework for accelerating CFD simulations by warm-starting. The method uses disturbance-aware partitioning, adaptive sampling, and region specific regression to model 3D flow fields. In addition, it uses geometry and flow conditions through multistage modulation and log-space regression to enforce positivity for flow variables. Experiments on 3 aircraft dataset show field prediction compared with 5 neural surrogate baselines. According to the results, DAFL warm-starts 18 test cases successfully and reports average CFD iteration speedups across 3 datasets.

Critical strengths
  1. The paper addresses a practical and important problem in scientific computing. Unlike standalone neural surrogates, the propsed approach uses learned predictions to initialize a numerical CFD solver, retaining the conventional solution process.
  2. The experimental results are promising. DAFL improves flow field prediction across three 3d datasets and achieves successful warm-starting in 18 test cases.
Critical weaknesses
  1. Limited methodological novelty. The proposed DAFL framework mainly combines existing techniques, such as adaptive sampling, region based regression, and log-space transformation. If simple idea can solve previously unsolved challenge, it is valuable. However, in this paper, the methodological contribution beyond their integration is not clear.
  2. Insufficient comparison with related warm-starting methods. The baselines are mostly general-purpose neural surrogates, which makes comparison unfair. Direct comparisons with existing CFD warm-starting approaches, such as Zhou et al. (https://www.sciencedirect.com/science/article/abs/pii/S0021999125001548?via%3Dihub) are missing, making it difficult to establish practical advantage of DAFL.
  3. Inadequate evaluation of solver acceleration and convergence. The reported speedups are only based on iteration counts, not actual computational time. Also, convergence is assessed through the drag coefficient, with no verification of equation residuals or final flow field accuracy.
  4. Question about generalization. The evaluation contains only 18 test cases, in which two datasets use fixed geometries and very narrow extrapolation settings. Therefore, the reported results do not fully support the claimed generalization across diverse geometries and flow conditions.
Questions for authors
  1. How does DAFL compare with exsiting learning-based CFD warm-starting approaches, particularly Zhou et al. (JCP 2025 from above)? Current baselines are all neural surrogates, making the comparisons not convincing
  2. Why does Algo 1 identify strong disturbances using GMM component with biggest signed mean? This may miss large negative velocity disturbances. How does it compare with a partition based on disturbance magnitude?
  3. Can authors separate the contributions of the following components: disturbance partition, adaptive sampling and region specific regression? Current ablation study removes all three together, making it hard to identify the source of improvement.
  4. What is the actual end 2 end wall-clock speedup, including neural inference and preprocessing? Do the warm-started solutions achieve similar eqauation residuals and final flow field to the cold-started solutions?
  5. How well does DAFL generalize to genuinely unseen geometries and meshes? Currently, 2 datasets use fixed geometry and X38 extrapolation tests only nearby angles of attack.
1 comment
AuthorsOpenReview ↗

Author Response to Reviewer YD1E Q2

Response to Q2

Q2: “Why does Algo 1 identify strong disturbances using GMM component with biggest signed mean? This may miss large negative velocity disturbances. How does it compare with a partition based on disturbance magnitude?”

We thank the reviewer for pointing out this ambiguity. Our implementation uses partitioning based on the absolute values of componentwise velocity deviations. Before fitting the GMM, we compute the absolute deviation of each velocity component from its corresponding background velocity. The GMM component means therefore reflect disturbance magnitudes rather than signed velocity deviations. For each velocity component, positive and negative deviations of equal magnitude produce identical input features, so large negative deviations are not underestimated because of their sign. The reported experimental results use this magnitude-based partitioning. The revised operation is described in Appendix B.1 (Binary Disturbance Label Generation), p. 15, lines 803 to 809, and p. 16, lines 810 to 825. Algorithm 1 appears on p. 16, lines 816 to 825, with the absolute-value operation on lines 820 to 821.

The original manuscript did not explicitly describe this absolute-value operation, which caused the ambiguity. The revised Algorithm 1 and corresponding description explicitly include this operation, aligning the stated partitioning criterion with the implementation used in the experiments.

Reviewer_7AVDOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DAFL (Disturbance-aware Field Learning), a learning-based framework for CFD warm-starting that aims to accelerate numerical simulations by predicting high-quality initial flow fields. The method adaptively partitions the flow domain into weak-disturbance, near-wall strong-disturbance, and outer strong-disturbance regions, employing region-specific sampling strategies and regression heads to improve the prediction of critical flow structures. It further incorporates multistage conditioning on geometry and flow conditions, logarithmic wall-distance features, and log-space regression to improve prediction accuracy and preserve the positivity of physical variables. Experiments on three complex 3D aerodynamic datasets demonstrate improved flow field prediction accuracy over existing baselines, successful CFD convergence, and reductions in solver iterations.

Critical strengths
  1. Insightful and application-oriented method design. The proposed method demonstrates a good understanding of the practical challenges in CFD warm-starting. Rather than simply applying an existing neural operator to predict initial flow fields, the authors identify several important characteristics of complex aerodynamic flows, such as spatially heterogeneous disturbances, steep near-wall gradients, and highly imbalanced flow variable distributions. Based on these insights, they develop a tailored framework with disturbance-aware sampling and regression, logarithmic wall-distance features, and log-space regression. These designs are well motivated by the specific requirements of CFD warm-starting.

  2. Comprehensive experiments and ablation studies. The paper conducts experiments on three complex 3D aerodynamic datasets covering different geometries and flow conditions, including both interpolation and extrapolation settings. It also provides extensive ablation studies to investigate the contributions of the proposed components. These experiments help demonstrate the effectiveness of the method and provide useful insights into the roles of different design choices.

  3. Rich and informative presentation of experimental results. The paper presents a wide range of experimental results, including quantitative comparisons, flow field visualizations, and CFD warm-starting performance evaluations.

Critical weaknesses
  1. Lack of comparisons with existing CFD warm-starting methods. Although the paper compares DAFL against five neural operator baselines, these methods are primarily designed for flow field prediction rather than CFD warm-starting. Several closely related works specifically focusing on CFD initialization and solver acceleration, such as Zhang et al. (2024) and Stumpo et al. (2026), are discussed in the Related Work section but are not included in the experimental comparisons. Since CFD warm-starting is the primary application and a central contribution of this paper, comparisons with existing warm-starting approaches, where feasible, would provide stronger evidence of the effectiveness and advantages of the proposed method.

  2. Insufficient analysis of baseline convergence failures. One of the main experimental findings is that DAFL achieves convergence in all test cases, while many baseline methods fail to converge, particularly on the X38 dataset. However, the paper does not provide a sufficiently detailed explanation of these failures. It remains unclear whether the failures are caused by physically invalid predictions (e.g., negative density or pressure), numerical instability, or simply the inability to satisfy the convergence criterion within the maximum number of iterations. More detailed analyses, such as solver residual curves, drag-coefficient convergence histories, and statistics on physically invalid predictions, would help clarify the causes of these failures and make the claimed robustness improvements more convincing.

  3. Insufficient justification and sensitivity analysis of the convergence criterion. The paper adopts an absolute drag-coefficient tolerance of 2.5×10−42.5\times10^{-4} to determine CFD convergence, but the rationale behind this specific choice is not sufficiently explained. Since convergence rates and solver speedups are central evaluation metrics, the conclusions may be sensitive to the selected tolerance. I suggest evaluating multiple convergence tolerances and reporting how the convergence success rates and speedup ratios change accordingly. In particular, a plot showing convergence success rate or solver iterations versus convergence tolerance for DAFL and the baselines would provide a clearer and more comprehensive comparison of their convergence behavior.

Questions for authors

See above.

Reviewer_1utqOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3
Presentation
1
Soundness
2
Read review
Paper summary

Disturbance-aware Field Learning (DAFL) proposes a data-driven framework for warm-starting solvers. DAFL succefully warmstarts and reduces the step needed for solver to converge on three different autho-generated datasets while other approaches dont converge in a consistent manner. DAFL is based on Point Transformer (PTv3) and adds on top of it calssifier for distubrance regions (targets based on GMM). The same GMM regions are also used for a disturbance-aware sampling scheme, where each region is voxelized at its own resolution during training. DAFL uses a second PTv3 with multi regression heads for the three defined regions (near, inner, and outer). Both backbones include multistage modulation based on the geometry and flow, and the model also predicts in log space and takes log wall distance as input. The paper includes several adapted baselines for the task including transolver, AB-UPT, and AdaField.

Critical strengths
  • The problem studied is of practical value but underexplored in the literature
  • The design choices are well-justified specially the disturbance aware sampling and regression, as shown in the ablation
  • Evaluation includes both surrogate-only results and solver-in-the-loop evaluation
  • The work targets 3D RANS, and the three datasets are designed with purpose
Critical weaknesses
  • Lacking details on the training of baselines (epochs, lr, ..etc). Tables 6 shows details of main models and ablations but does not include the baselines (correct me if this was misunderstood)
  • log regression of turbulence variables is feature engineering and should be consistenly done in all baselines since they were adapted by the authors, and it can be a major cause of why other approaches fail to warm-start the solver in a consistent manner.
  • FLOPS, GPU memory, and other complexity measures were not reported.
  • The impact of Disutrbance aware sampling and regression are coupled, an ablation should seperate the impact of both
Questions for authors
  • Are you predicting on the same mesh for warm-starting all methods?
  • Clarify if the reported 16M params count is for each backbone or includes both
  • Can you inlcude AB-UPT or AdaField with log wall distance input and log regression on the X38 dataset (warm-starting)? This is a decisive experiment since all models fail there.
  • Can you clarify why only AB-UPT doesn't take wall distance as input? or is this only an issue in the text?
  • Can you seperate the impact of Disutrbance aware sampling and regression?
  • Can you try the sampling approach on another model? This can have an impact broader than a single framework
  • Did you do any lr sweep for the baselines?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.