FIGConvNet: Factorized Global Convolution

• FIGConvNet is a deep-learning surrogate model that employs factorized implicit global convolution to predict spatial aerodynamic fields on arbitrary vehicle geometries.
• It uses a U-shaped encoder–decoder architecture with global feature mixing that scales quadratically, enabling efficient inference on large point clouds.
• Evaluation on the DrivAerML dataset shows high regression accuracy for integrated drag and lift metrics, robust generalization, and competitive performance compared to volumetric and graph-based methods.

FIGConvNet (Factorized Implicit Global Convolution Network) is a specialized deep-learning surrogate model for predicting surface aerodynamic quantities on arbitrary vehicle geometries. Developed for exterior vehicle aerodynamics, it efficiently infers spatially distributed surface pressure and wall-shear stress (WSS) fields from CAD-derived geometries, achieving high fidelity at O(N²) computational cost for N surface points—substantially improving scalability over cubic-cost alternatives such as volumetric 3D convolution or graph neural network (GNN) layers. Its architecture leverages a factorized implicit global convolution mechanism, enabling global information exchange across large surface meshes and delivering state-of-the-art predictive accuracy for integral aerodynamic coefficients such as drag and lift within the open-source NVIDIA PhysicsNeMo-CFD framework (Tangsali et al., 14 Jul 2025).

1. Model Architecture and Theoretical Formulation

FIGConvNet ingests a uniformly sampled point cloud from the vehicle's triangulated STL surface mesh, typically comprising 500,000 points, each annotated with 3D coordinates and estimated surface normals. Instead of constructing an explicit 3D volumetric grid, the model adopts a factorized implicit representation, parameterizing the 3D problem domain as a stack of mutually orthogonal 2D "slices." This enables global feature communication with quadratic, rather than cubic, scaling behavior.

The network follows a U-shaped encoder–decoder paradigm:

• Encoder: Multi-stage down-sampling modules aggregate local and mid-range contextual information over the point cloud.
• Decoder: Up-sampling layers reconstruct high-resolution output fields, with skip connections preserving geometric detail.

At each stage, a global convolution operator K is applied instead of conventional 3D convolutions. The per-point feature update at layer ℓ+1 is given by:

$$H^{(\ell+1)} = \sigma\left( \operatorname{Conv2D}_b\left( \operatorname{Conv2D}_a\left( H^{(\ell)}; W_a \right); W_b \right) + b \right)$$

where $\operatorname{Conv2D}_a$ and $\operatorname{Conv2D}_b$ act along complementary 2D parameterizations (2D implicit grids obtained by projecting the 3D surface), and $W_a, W_b$ are learnable weight tensors. This factorization produces a global mixing of features that scales as O(N²) instead of O(N³), permitting practical inference at large sample counts. The global mixing kernel $K$ is implemented as a learnable function over pairs of 2D-embedded point coordinates $\phi(x_i), \phi(x_j)$.

2. Training Protocol and Dataset Characteristics

The model is trained and validated on the DrivAerML dataset, encompassing 500 parametric shape variants of the DrivAer notchback automotive geometry. Hybrid RANS–LES simulations generate surface cell-centered fields for pressure and wall-shear stress vectors ($\tau_x, \tau_y, \tau_z$), as well as derived integral aerodynamic coefficients ($C_D$, $C_L$).

Preprocessing includes uniform sampling of 500,000 points on each CAD mesh and interpolation of CFD solutions onto these points, accompanied by per-point normal computation. The training/validation split assigns 436 geometries for training and 48 for validation, with the hold-out set constructed to test generalization, explicitly including both high-drag and low-drag extremes along with random samples.

Optimization uses AdamW (weight decay $1\mathrm{e}{-5}$), initial learning rate $\operatorname{Conv2D}_a$0 (cosine decay), 200 epochs, and batch size of one geometry (entire point cloud). The loss combines per-point MSE with an aerodynamically integral term:

$\operatorname{Conv2D}_a$1

with $\operatorname{Conv2D}_a$2 chosen as $\operatorname{Conv2D}_a$3 to moderate the trade-off between local field accuracy and integral force prediction.

3. Evaluation Metrics and Benchmarking Results

Within the NVIDIA PhysicsNeMo-CFD framework, FIGConvNet's performance is quantified using both pointwise and integral metrics:

• Surface L₂ error and area-weighted L₂ error for pressure and all WSS components, evaluated after remapping predictions to CFD surface cell centers.
• Coefficient of determination (R²) quantifies regression accuracy for predicted drag and lift coefficients.
• Spearman rank correlation (ρ) and mean/worst-case absolute percent deviation capture trend fidelity for ranking by aerodynamic coefficients.

Key results on the 48-sample hold-out validation set include:

• L₂ errors (surface): pressure: 0.21; WSSₓ: 0.32, WSSᵧ: 0.62, WSS_z: 0.53
• Area-weighted L₂ errors: pressure: 0.14; WSSₓ: 0.16, WSSᵧ: 0.34, WSS_z: 0.29
• Integral metrics:
  • Drag: $\operatorname{Conv2D}_a$4, $\operatorname{Conv2D}_a$5, mean $\operatorname{Conv2D}_a$6, max $\operatorname{Conv2D}_a$7
  • Lift: $\operatorname{Conv2D}_a$8, $\operatorname{Conv2D}_a$9, mean $\operatorname{Conv2D}_b$0, max $\operatorname{Conv2D}_b$1

On a denser 10M-point cloud, the model retains $\operatorname{Conv2D}_b$2 and $\operatorname{Conv2D}_b$3, indicating scalability of predictions.

Comparative benchmarking:

• X-MeshGraphNet: FIGConvNet exhibits 50–100% higher WSS errors but surpasses X-MeshGraphNet in drag/lift regression due to its explicit integral loss term.
• DoMINO: DoMINO achieves lower pointwise/area-weighted errors and marginally higher $\operatorname{Conv2D}_b$4, but requires coupling to a more computationally expensive volumetric solver.

4. Computational Scaling and Generalization Properties

FIGConvNet's core innovation is the factorized global convolution, yielding O(N²) runtime complexity per layer. Inference on 500,000-point clouds completes in 5–10 seconds on a single NVIDIA V100 GPU, a substantial improvement over the 30–50 second runtime of cubic-scaling graph-neural or volumetric convolution methods.

The model operates on point clouds without reliance on explicit mesh connectivity, conferring robustness to variations in CAD triangulation; only the density of point sampling influences performance. The chosen validation split—encompassing extreme geometries—demonstrates that the model generalizes effectively, maintaining $\operatorname{Conv2D}_b$5 and $\operatorname{Conv2D}_b$6 even for outliers within the DrivAerML parameter space.

5. Strengths, Limitations, and Prospects for Enhancement

Strengths include quadratic runtime complexity while capturing global aerodynamic interactions, integral coefficient loss regularization for high-fidelity drag/lift regression, and a mesh-agnostic point cloud framework obviating the need for mesh standardization.

Limitations:

• Current implementation predicts only surface aerodynamic fields; volume-field prediction is not supported, though indicated as an area of ongoing research.
• Lacks explicit enforcement of physical laws such as mass or momentum conservation; the approach is entirely data-driven.
• Achieving optimal balance between local accuracy and integral force fidelity requires careful hyperparameter optimization, especially the weighting of loss terms.

Potential improvements outlined by the authors:

• Extension of the factorized convolution strategy to volumetric (3D) architectures to support full-field predictions with O(N²) or O(N log N) cost.
• Integration of physics-informed loss functions reflecting residuals for continuity and momentum, enhancing consistency in regions of complex flow.
• Addition of uncertainty quantification (via, e.g., Monte Carlo dropout or deep ensembles) to constrain error bounds in safety-critical design.
• Seamless coupling with conventional CFD solvers for rapid solver initialization (warm starts), thereby accelerating hybrid simulation workflows.

6. Application Scope and Comparative Summary

FIGConvNet is intended for accelerating the evaluation phase in data-driven, high-fidelity vehicle aerodynamic design, supporting rapid comparison of candidate geometries and integration into Computer-Aided Engineering (CAE) pipelines. The model's architecture and benchmarked performance underpin its utility in scenarios requiring both granular field predictions and reliable integrated force estimates across a diverse geometric design space.

In summary, FIGConvNet represents a leading solution for point-cloud-based aerodynamic surrogate modeling, balancing global receptive field, mesh independence, computational efficiency, and drag/lift regression accuracy. Remaining research directions focus on volume-field support and enforcing stronger physical constraints during training for enhanced deployment in design-critical and safety-regulated automotive environments (Tangsali et al., 14 Jul 2025).