Source-linked AI summary
FSAN: Flow State Attention Network for Aerodynamic Prediction
Wenxuan Jin, Jianguo Yao, Haibing Guan, Xijun Li
TL;DR
Aerodynamic CFD is expensive, and existing neural surrogates have limited evidence under complex flow conditions and spatially uniform conditioning. FSAN separately encodes geometry and flow, partitions geometry into learnable latent flow states, and couples them through state-specific attention. Across two public benchmarks, it achieves the best accuracy among compared methods with fewer parameters, but at higher computational cost.
Problem
Existing aerodynamic surrogates are evaluated over narrow flow-condition ranges and inject global flow information uniformly across surface points, while CFD remains expensive for iterative and real-time use.
Method
FSAN separately encodes point clouds and flow conditions, adaptively partitions geometry into multiple latent flow states, and uses flow state attention for fine-grained geometry–flow interaction.
Results
FSAN achieves the best accuracy among compared methods on Emmi-Wing and DrivAerNet++ while using fewer parameters.
Takeaways & Limitations
FSAN is a promising neural surrogate for aerodynamic prediction on public benchmarks with diverse flow conditions and geometries.
Takeaways & Limitations
Cross-domain generalization, data-scarce performance, and out-of-distribution behavior remain unexplored, while FSAN has higher computational cost than lightweight baselines.
Abstract
from arXiv · showhide
Accurate aerodynamic prediction is critical for designing fuel-efficient and safe transportation systems such as aircraft and automobiles, yet traditional computational fluid dynamics (CFD) simulations remain computationally expensive and expertise-intensive, severely limiting their use in iterative design and real-time analysis. Existing deep learning surrogates suffer from two major limitations: (i) they are evaluated on datasets with narrow flow-condition ranges, leaving their performance under complex flow conditions undemonstrated; (ii) they treat global flow conditions as a single vector injected uniformly across all surface points, ignoring that different geometric regions experience distinct local flow phenomena, which degrades prediction accuracy under complex flow conditions. To address these limitations, we propose the Flow State Attention Network (FSAN). FSAN separately encodes point cloud and flow conditions, then partitions the geometry into multiple flow states via learnable soft assignments, and uses flow features to update these state representations, which in turn influence point cloud features through state changes. This enables fine-grained, state-specific interaction between geometry and flow information. Extensive experiments on two well-recognized aerodynamic benchmarks demonstrate that FSAN achieves the highest accuracy among the methods compared in this work at a higher computational cost. On Emmi-Wing, FSAN reduces the Relative L2 (REL-L2) error by over 20\% compared to the strongest baseline (Transolver), and on DrivAerNet++, it achieves a 10\% reduction compared to the strongest baseline (AdaField). These results establish FSAN as a promising neural surrogate on public benchmarks with diverse flow conditions and geometries.
1 INTRODUCTION
Traditional CFD is too costly for iterative and real-time aerodynamic analysis, while existing neural surrogates are insufficiently tested under diverse flow conditions and use spatially uniform conditioning. FSAN addresses these gaps through flow-state partitioning and state-specific geometry–flow interaction, achieving stronger benchmark accuracy with fewer parameters.
- CFD simulations require hours to days per configuration and substantial hardware resources, limiting iterative design optimization and real-time analysis.
- Existing neural surrogates are often evaluated over narrow flow-condition ranges and lack demonstrated performance under complex conditions.
- FSAN separately encodes geometry and flow conditions, then adaptively partitions the point cloud into multiple latent flow states.
- Flow state attention enables fine-grained, state-specific interaction between geometry and flow information rather than uniform global conditioning.
- Over 20% lower REL-L2 error on Emmi-Wing versus Transolver accompanied 14% fewer parameters, while DrivAerNet++ showed 10% lower REL-L2 versus AdaField with 94% fewer parameters.
- FSAN produces spatially coherent latent partitions that are progressively refined by flow information without explicit physical supervision.
2 RELATED WORK
Related aerodynamic surrogates capture geometric variation effectively, but commonly use narrow flow-condition ranges and spatially uniform conditioning. These limitations motivate representations that can capture region-dependent responses to local flow phenomena.
- Many aerodynamic surrogates emphasize geometric features while commonly restricting flow conditions to narrow ranges.
- FiLM and cross-attention condition point features uniformly within each sample, exposing all points to shared flow information.
- Spatially uniform conditioning does not explicitly represent region-dependent responses arising from distinct local flow phenomena.
3 METHOD
FSAN predicts per-point pressure coefficients by separately encoding geometry and flow conditions, then progressively refining geometry-driven flow states through state-specific cross-attention. Its stacked architecture combines local point interactions, flow-token encoding, and projection of flow-conditioned states back to individual points.
- Inputs and outputs: FSAN takes a point cloud and global flow conditions as input and outputs a predicted pressure coefficient for every surface point.For aircraft, the flow vector includes angle of attack, Mach number, and Reynolds number.
- Network architecture: Each of L stacked blocks combines local attention, a flow condition encoder, and a flow state attention module.Local attention uses neighboring points, while the other modules process global flow information and state-specific interactions.
- Local attention: The local attention layer updates each point by attending to its k nearest neighbors and then applies residual and feed-forward processing.Point coordinates remain fixed across layers and provide relative-position information for local attention.
- Flow condition encoding: The flow condition encoder maps the global flow vector into K flow tokens, each with hidden dimension C, for subsequent cross-attention.For aircraft tasks, the encoded flow vector represents angle of attack, Mach number, and Reynolds number.
- Flow state attention: Flow state attention assigns points softly to multiple states, aggregates state features, updates each state using flow-token cross-attention, and projects refined states back to points.The assignments are geometry-driven in the first block and become jointly determined by geometry and evolving flow-aware features in later blocks.
- Design distinction: Unlike Transolver, FSAN separates geometry and flow streams, progressively refines state assignments, and lets different states extract different information from shared flow tokens.Projecting the refined states back to points gives differently assigned regions distinct conditioning signals under identical inflow parameters.
4 EXPERIMENTS
Experiments evaluate FSAN against strong baselines on two public aerodynamic datasets, measuring predictive accuracy, computational cost, conditioning mechanisms, and model-size effects. FSAN achieves the best reported accuracy, while its inference cost is higher than Transolver and its larger variants converge better.
- Evaluation setup: Experiments compare FSAN with PointTransformerV3, Transolver, and AdaField on Emmi-Wing and DrivAerNet++.Emmi-Wing varies multiple flow parameters, while DrivAerNet++ provides broad geometric diversity and is augmented with variable velocity.
- Main results: FSAN consistently outperforms all baselines across four metrics on both datasets.On Emmi-Wing, REL-L2 is 2.59% versus Transolver’s 3.24%; on DrivAerNet++, it is 17.91% versus AdaField’s 19.89%.
- Main results: FSAN reduces REL-L2 by over 20% on Emmi-Wing and 10% on DrivAerNet++ while using 14% and 94% fewer parameters, respectively.The comparisons are against Transolver on Emmi-Wing and AdaField on DrivAerNet++.
- Computational cost: FSAN’s inference is about 4× slower than Transolver and uses about 5.8× more GPU memory.The overhead primarily comes from non-hierarchical local attention retaining full-resolution activations for all 8,192 points.
- Mechanism isolation: The flow state attention mechanism achieves the best accuracy among the tested conditioning methods in controlled Emmi-Wing ablations.The ablations replace only the geometry–flow fusion mechanism while keeping the backbone, local attention, and flow encoder fixed.
- Model size: Both final error and convergence speed improve with model size, with the 16M-parameter configuration achieving the lowest tested error.The selected configuration uses 12 blocks and feature dimension 256.
5 DISCUSSION
The discussion identifies unresolved scope and efficiency limitations: cross-domain generalization, data-scarce performance, and out-of-distribution behavior remain untested, while FSAN is computationally heavier than lightweight baselines. Proposed future directions target these gaps through transfer, data-efficient, OOD, and efficient-attention methods.
- Scope limitations: Cross-domain generalization remains unexplored because FSAN is evaluated separately on Emmi-Wing and DrivAerNet++.The authors propose joint training and transfer learning as future directions.
- Scope limitations: FSAN’s performance in data-scarce engineering scenarios has not been investigated.Future work proposes meta-learning and multi-fidelity approaches for this setting.
- Scope limitations: Out-of-distribution evaluation is missing, partly because suitable public aerodynamic datasets are scarce.The authors plan systematic OOD evaluation when appropriate datasets become available.
- Efficiency limitation: FSAN incurs higher computational cost than lightweight baselines such as Transolver because its non-hierarchical design retains full-resolution activations.Proposed remedies include hierarchical approximations and efficient attention variants.
- Future work: Uncertainty quantification remains an additional future direction for safety-critical aerodynamic prediction.The authors identify it separately from the computational and evaluation limitations.
6 CONCLUSION
The conclusion presents FSAN as a flow-conditioned, state-specific surrogate that addresses narrow flow-condition evaluation and uniform conditioning. Across two public benchmarks, it reports the best accuracy with fewer parameters but higher computational cost, while the study provides code and uses public datasets.
- Contribution: FSAN addresses narrow flow-condition evaluation and homogeneous conditioning by learning a soft, flow-conditioned partition into latent states.The state-specific interaction links geometry and flow conditions at a finer granularity.
- Conclusion: FSAN achieves the best accuracy among the compared methods on two public aerodynamic benchmarks while using the fewest parameters.The conclusion reports higher computational cost as the trade-off.
- Conclusion: FSAN reduces REL-L2 error by 20% on Emmi-Wing with 14% fewer parameters and by 10% on DrivAerNet++ with 94% fewer parameters.These comparisons summarize the reported benchmark results.
- Reproducibility: The study uses public Emmi-Wing and DrivAerNet++ datasets and provides anonymous source code with training and evaluation scripts.Baseline implementations are not redistributed because they are publicly available under their original licenses.
A.3 DATASET SELECTION RATIONALE
The study selects datasets spanning complementary aerodynamic variation and compares several adapted baseline architectures under a consistent conditioning protocol.
- Dataset selection rationale: Emmi-Wing provides multi-parameter flow variation, while DrivAerNet++ provides broad geometric diversity under fixed velocity.PIDA augmentation adds variable velocity to DrivAerNet++, increasing flow-condition complexity.
- Baseline selection rationale: The baseline set includes PointTransformerV3, Transolver, and AdaField, with architectural adaptations described for aerodynamic pressure prediction.Transolver is enlarged to approximately 18M parameters for capacity-matched comparison; AdaField has approximately 250M parameters.
- Baseline selection rationale: The comparison excludes methods lacking native variable-flow support, requiring substantial modifications, or lacking publicly available code.AB-UPT is reproduced separately under a comparable protocol in Appendix E.1.
- Baseline selection rationale: Models without native flow encoders receive flow parameters concatenated with point coordinates as a 6-dimensional input.This protocol is applied consistently to geometry-only baselines, although the authors note that concatenation may cause partial misinterpretation of flow parameters as spatial dimensions.
C EVALUATION METRICS
The evaluation uses absolute and scale-normalized error metrics to assess aerodynamic pressure predictions across samples with different pressure magnitudes.
- Evaluation metrics: MSE and MAE measure absolute prediction errors, with MSE more sensitive to large deviations because errors are squared.These metrics quantify unnormalized prediction discrepancies.
- Evaluation metrics: REL-L1 and REL-L2 are scale-normalized relative errors that enable fair comparisons across samples with different pressure magnitudes.Their normalization makes relative performance comparable when pressure scales differ.
D TRAINING SETTINGS
All models are trained under matched optimization, hardware, batching, and evaluation settings, with results averaged across repeated runs.
- Training settings: All models use identical hyperparameters and hardware to support fair comparison.Training uses AdamW, cosine annealing, and 50 epochs with warmup iterations.
- Training settings: Training uses distributed execution on 8 NVIDIA RTX 4080 SUPER GPUs with a global batch size of 8.Each GPU uses batch size 1 under PyTorch distributed data parallel.
- Training settings: The loss function is mean absolute error, and test samples are uniformly downsampled to 8,192 points.The same point-sampling protocol is used during training and testing.
- Training settings: Reported means and standard deviations are computed from three runs with different random seeds.This repeated-seed protocol quantifies variation across training runs.
E.1 COMPARISON WITH AB-UPT
The AB-UPT comparison uses a restricted Emmi-Wing subset because AB-UPT requires substantial volume-field data, while additional analyses examine flow-state count and conditioning strategy.
- Comparison with AB-UPT: AB-UPT is evaluated on the first 1,000 Emmi-Wing samples because its volume-field input requires approximately 5 TB of additional data.The subset contains 800 training samples and 200 testing samples, using normalized surface pressure.
- Comparison with AB-UPT: FSAN performs slightly better than AB-UPT while using only surface information.The authors therefore report a conservative conclusion of significantly lower training-data cost for comparable performance.
- Flow-state analysis: Accuracy improves as the number of flow states increases from 1 to 8, then degrades at M = 12 and M = 16.The authors attribute the later degradation to possible over-partitioning and use M = 8 in the full model.
- Conditioning strategy comparison: Concatenation achieves the best accuracy among concatenation, FiLM, and AdaLN-Zero on Transolver for Emmi-Wing.This result supports concatenation as the baseline adaptation protocol for geometry-only models without native multi-parameter flow inputs.
E.4 COMPUTATIONAL COST ANALYSIS
FSAN’s computational cost is assessed across hardware, point densities, and modules. Its main overhead comes from the non-hierarchical local-attention backbone, while accuracy is best at the training density of 8,192 points.
- Cost settings: FSAN’s DrivAerNet++ cost is measured on 8× NVIDIA RTX 4080 SUPER GPUs with batch size 1 and 8,192 points per sample.The point-cloud processing modules remain identical to those used on Emmi-Wing.
- Module-level breakdown: The local attention layer dominates FSAN’s inference time and peak memory, while flow state attention accounts for less than 3.2% of inference time.The module-level analysis attributes the overhead primarily to retaining full-resolution activations for all 8,192 points.
- Sampling-density sensitivity: FSAN achieves its best accuracy at 8,192 sampled points, matching the training density.Accuracy decreases noticeably below 4,096 points and slightly decreases above 8,192 points, attributed to training–testing density mismatch.
E.6 CONSISTENCY BETWEEN FULL-SURFACE AND SAMPLED PREDICTIONS
Full-surface inference is consistent with 8,192-point sampled inference, while qualitative analyses show that FSAN’s predictions and learned flow-state partitions remain meaningful across representative flow conditions. The partitions evolve across blocks as flow information is incorporated and show associations with shock regions without implying fixed physical regimes.
- Prediction consistency: Full-surface and sampled predictions are nearly identical, with REL-L2 differing only marginally at 2.63% versus 2.59%.The full surface is randomly chunked into 8,192-point batches, processed independently, and stitched back to the original point order.
- Prediction quality: FSAN achieves visibly lower errors than the baselines on moderate, high-angle-of-attack, and high-Mach/Reynolds-number samples.The representative cases use signed-error visualizations saturated at ±0.015, with unclipped statistics reported separately.
- Prediction quality: Under the extreme combined condition, all methods exhibit larger errors, suggesting that the difficulty reflects the flow regime’s intrinsic complexity rather than a specific FSAN limitation.The extreme case combines high angle of attack, Mach number, and Reynolds number.
- Flow-state evolution: Flow-state partitions are initially similar across samples, then diverge across deeper blocks as cross-attention updates incorporate flow conditions.Their dominant orientation shifts from spanwise in early blocks to chordwise in later blocks, and state indices are independent across blocks.
- Aerodynamic-structure association: One flow state shows higher activation on shock regions than non-shock regions, while other states avoid shocks.Because assignments are softmax-based, this association does not imply a hard one-to-one mapping to physical flow regimes.