Organizations: Tsinghua University · Peking University · Nanyang Technological University · Shanghai Jiao Tong University · The Hong Kong University of Science and Technology
Neural operators provide efficient surrogates for computational fluid dynamics (CFD), but most are designed for static geometries and are not equipped to model unsteady flows induced by continuously moving boundaries. We study dynamic-boundary 4D CFD prediction for articulated moving geometries and introduce RoboCFD-4D, a temporally resolved robotic CFD benchmark spanning diverse morphologies and flow conditions. To address the computational and autoregressive challenges of this setting, we propose DynaSolver, a neural operator that explicitly conditions on boundary evolution and causal flow history. DynaSolver employs an efficient physical-token formulation that amortizes point-to-token projection across the operator stack, substantially reducing repeated point-space spatio-temporal computation while retaining accurate point-level reconstruction. We further introduce Replay-Augmented Parallel Training, which exposes the model to autoregressively generated histories while preserving parallel dense supervision, narrowing the gap between ground-truth-conditioned training and autoregressive rollout. Experiments on RoboCFD-4D show that DynaSolver achieves the best full-rollout accuracy among a broad range of learned CFD surrogates, with its advantage increasing as rollout proceeds, indicating substantially reduced autoregressive error accumulation. Project page: https://made-gpt.github.io/DynaSolver-ProjectPage.
Figures & tables
Dataset
Domain
Dim.
Boundary Dynamics
Sequences
CFD States
AirfRANS ( Bonnet et al., 2022 )
Airfoil aerodynamics
2D
Static solid
N/A
1,000
CFDBench ( Luo et al., 2023 )
General fluid dynamics
2D+t
Static solid
739
301,671
MegaFlow2D ( Xu et al., 2023 )
External / internal flow
2D+t
Static solid
∼ 3,000
> 2M
RealPDEBench ( Hu et al., 2026a )
real scenarios
2D+t
Static/Moving solid
736
∼ 2.6M
Eagle ( Janny et al., 2023 )
Drone
2D+t
Moving Source, Static solid
1184
∼ 1M
FlowBench ( Tali et al., 2025 )
Complex-geometry flow
2D+t
Static solid
1,150
278,300
Table 1: Comparison of representative CFD learning datasets across spatial dimensionality, temporal resolution, and boundary dynamics. “Dim.” denotes spatial and temporal dimensionality, where 3D+t indicates temporally resolved three-dimensional CFD.
Figure 2: Overview of DynaSolver. Given the spatial condition, boundary evolution, and previous CFD states over a causal window, DynaSolver lifts and fuses the point-wise inputs before mapping them into separate surface and volume physical tokens. A single Fast Slice/Deslice pair encloses the spatio-temporal operator stack, where self-attention, cross-attention, and causal attention jointly model spatial and temporal flow interactions. The reconstructed point features are decoded into a residual flow field and added to the previous CFD state to predict the current velocity and pressure.
Figure 3: Comparison of training strategies for autoregressive flow prediction. Dense Unrolled Training (a) and Dense Pushforward Training (b) construct autoregressive histories sequentially, whereas Parallel Teacher-Forcing Training (c) predicts all transitions in a clip simultaneously from ground-truth histories. Our Replay-Augmented Parallel Training (d) decouples autoregressive history generation from student optimization: a frozen model first rolls out replay inputs (d.1), which are then used for parallel dense supervision of the current model (d.2). Red solid arrows → denote feedforward computation, while green dashed arrows ⇢ denote backpropagation.
Method
Per-frame Velocity ↓
Per-frame Pressure ↓
Avg. Vel. ↓
Avg. Pre. ↓
Overall ↓
1
6
12
18
24
29
1
6
12
18
24
29
GNOT ( Hao et al., 2023 )
11.38
21.14
28.47
34.58
40.01
43.99
34.08
72.89
93.49
98.66
102.21
99.18
30.56
87.84
31.55
GAOT ( Wen et al., 2025 )
14.46
19.54
23.58
26.93
29.66
31.19
40.96
57.70
66.52
75.18
84.08
90.67
24.60
70.25
25.32
CNO ( Raonic et al., 2023 )
9.32
19.36
24.27
28.49
35.25
47.04
34.24
77.09
87.29
98.14
108.89
120.71
27.12
91.52
28.28
P3D ( Holzschuh et al., 2026 )
9.32
17.07
22.62
26.70
30.44
32.36
32.55
62.66
76.03
85.25
96.86
102.08
23.72
78.82
24.61
MGN ( Pfaff et al., 2021 )
9.49
19.70
23.69
25.88
27.66
28.55
29.32
65.02
79.25
86.55
96.78
99.86
23.48
79.44
24.51
Table 2: Quantitative comparison across rollout horizons on RoboCFD-4D Base. We report the relative L2 errors (%) of velocity and pressure at selected rollout steps, together with their full-rollout averages and the overall error. The best results among learned methods are highlighted in bold , while the second-best are underlined . Persistence is excluded from ranking.
Figure 4: Rollout error evolution on RoboCFD-4D Base. Relative L2 errors are reported across the full autoregressive rollout for the overall flow field, velocity, and pressure.
Figure 5: Qualitative comparison across autoregressive rollout. Ground-truth fields and corresponding error maps are shown at selected rollout steps for an unseen robot morphology. DynaSolver shows lower long-horizon errors than the baselines, particularly for velocity.
Figure 6: Training cost versus overall rollout error. Overall rollout relative L2 error is plotted against training time; DynaSolver achieves the lowest rollout error while maintaining low training cost.
Variants
Average Velocity L2↓
Average Pressure L2↓
Overall L2↓
Train Time (h) ↓
Inference Time (ms) ↓
Params
Transolver
20.12
71.39
21.06
1.52
8.09
3.87M
+ Boundary Evolution
19.83
66.45
20.69
1.53
8.14
3.87M
+ Causal Temporal Attention
17.53
45.91
17.98
6.68
14.79
5.98M
+ Surface-Volume Dual Branch
16.24
42.97
16.68
11.47
27.34
8.09M
+ Amortized Token-Space Operator
18.39
53.13
18.98
3.09
9.99
8.09M
+ Pointwise MLP Decoder
15.65
44.30
16.14
2.31
10.85
8.62M
Table 3: Progressive design ablation on RoboCFD-4D Mini. We report the full-rollout relative L2 errors (%) together with training cost, inference latency, and model size. The blue row is ours.
Table 4: Training-strategy ablation on RoboCFD-4D Mini. Left: full-rollout relative L2 errors (%) and training cost under the same DynaSolver architecture, TF represents Teacher Forcing. Right: training convergence over epochs. The blue row is ours.
W
Avg. Vel. L2↓
Avg. Pre. L2↓
Overall L2↓
Train (h) ↓
1
18.27
66.17
19.22
0.98
2
16.10
48.52
16.67
1.06
3
16.04
48.11
16.60
1.53
4
15.65
44.30
16.14
2.31
5
15.79
45.48
16.29
3.07
6
15.75
45.89
16.28
3.21
Table 5: Sensitivity Analysis on RoboCFD-4D Mini. We vary the temporal window length W while fixing the slice count at S=32 (left), and vary S while fixing W=4 (right). The blue row is ours.
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Input
Feature
Dim.
Samples
Representation
ct
normalized position xt
3
S/V
dimensionless
signed distance ϕt
1
S/V
m
nearest outward normal nt
3
S/V
unit vector
inflow direction d∞
3
S/V
unit vector
normalized inflow speed U∞
1
S/V
dimensionless
bt−1→t
previous signed distance ϕt−1
1
S/V
m
Appendix
Table 6: Pointwise input features used by DynaSolver. S and V denote surface-associated and volume grid nodes, respectively.
Figure 7: Overview of the RoboCFD-4D data generation pipeline. Public robot assets are standardized and repaired at the link level, followed by prescribed motion synthesis and reusable link-local SDF construction. The transformed link-local fields define the prescribed moving immersed boundary for multi-condition CFD simulations. Aligned volumetric fields and tracked near-surface observations are then extracted and subjected to geometric, numerical, and dataset-level validation before inclusion in the final published dataset.
Speed (m/s)
Yaw
Condition ID
2
−20∘
u2_yaw_m20
2
0∘
u2_yaw_p0
2
20∘
u2_yaw_p20
3
−10∘
u3_yaw_m10
3
10∘
u3_yaw_p10
4
−20∘
u4_yaw_m20
Appendix
Table 7: Wind conditions and repository identifiers used in RoboCFD-4D.
Figure 8: Ground-truth flow visualization from RoboCFD-4D. Six articulated robots are shown at rollout frames 6, 12, 18, and 24 under a 2 m/s inflow and 0∘ yaw. Each pair of columns shows pressure and normalized velocity magnitude ∣U∣/U∞ around the moving robot. Row labels indicate the robot and whether its morphology is included in training or held out for evaluation.
Figure 9: Ground-truth flow visualization from RoboCFD-4D. Five articulated robots are shown at rollout frames 6, 12, 18, and 24 under a 2 m/s inflow and 0∘ yaw. Each pair of columns shows pressure and normalized velocity magnitude ∣U∣/U∞ around the moving robot. Row labels indicate the robot and whether its morphology is included in training or held out for evaluation.
Figure 10: Seen morphology under an unseen inflow condition. Error maps of four models on a quadruped, evaluated at 4 m/s and 0∘ yaw. The morphology belongs to a family that the model sees during training, whereas this inflow condition is held out, so the figure isolates the effect of the flow regime with the boundary geometry in distribution. Columns alternate between the pressure and the velocity field at rollout steps 6, 12, 18 and 24, and all panels share one colour scale.
Figure 11: Held-out morphology under an in-range inflow condition. The same layout as Figure 10 , for an additional humanoid body that is held out from training in all conditions, evaluated at 2 m/s and −20∘ yaw, which lies within the range of conditions the model was trained on. The inflow therefore remains in distribution and the figure isolates the effect of an unfamiliar morphology. Colour is shared by all panels of the figure.
Figure 12: Failure under stronger inflow conditions. Absolute pressure errors ∣ppred−pgt∣ at rollout step t=24 for a quadruped at 0∘ yaw under inflow speeds of 2 , 4 , 6 , and 8m/s . A shared color scale is used across all methods and conditions. While errors are relatively localized at low inflow speeds, they increase sharply and spread throughout the wake at higher speeds, indicating that strongly unsteady flow conditions remain challenging for current neural CFD surrogates.
School of Computer Science, Sichuan University, Chengdu, China · National Key Laboratory of Fundamental Algorithms and Models for Engineering Numerical Simulation, Chengdu, China