AB-UPT: Scaling Neural CFD Surrogates for High-Fidelity Automotive Aerodynamics Simulations via Anchored-Branched Universal Physics Transformers

TL;DR

This work tackles the challenge of scaling neural surrogates for high-fidelity automotive CFD, where industrial meshes exceed 100 million cells. It introduces AB-UPT, a three-branch Transformer-based neural operator with anchor attention that decouples geometry encoding from surface and volume predictions and enables linear-inference complexity via a small set of anchor tokens. A key innovation is enforcing physics consistency by constructing a divergence-free vorticity field as a hard constraint through a neural-field decoder, enabling accurate and physically plausible predictions. AB-UPT achieves state-of-the-art accuracy on surface and volume fields across large datasets, trains on a single GPU in under a day, and can predict from CAD geometry without meshing, illustrating practical impact for design cycles and CAD-to-CFD workflows. The work also provides extensive scaling analyses (problem, model, data) and practical guidance for deploying neural surrogates in industrial CFD contexts, with future directions in parallelization, uncertainty, and broader physics-informed extensions.

Abstract

Recent advances in neural surrogate modeling offer the potential for transformative innovations in applications such as automotive aerodynamics. Yet, industrial-scale problems often involve volumetric meshes with cell counts reaching 100 million, presenting major scalability challenges. Complex geometries further complicate modeling through intricate surface-volume interactions, while quantities such as vorticity are highly nonlinear and must satisfy strict divergence-free constraints. To address these requirements, we introduce AB-UPT as a novel modeling scheme for building neural surrogates for CFD simulations. AB-UPT is designed to: (i) decouple geometry encoding and prediction tasks via multi-branch operators; (ii) enable scalability to high-resolution outputs via neural simulation in a low-dimensional latent space, coupled with anchored neural field decoders to predict high-fidelity outputs; (iii) enforce physics consistency by a divergence-free formulation. We show that AB-UPT yields state-of-the-art predictive accuracy of surface and volume fields on automotive CFD simulations ranging from 33 thousand up to 150 million mesh cells. Furthermore, our anchored neural field architecture enables the enforcement of hard physical constraints on the physics predictions without degradation in performance, exemplified by modeling divergence-free vorticity fields. Notably, the proposed models can be trained on a single GPU in less than a day and predict industry-standard surface and volume fields within seconds. Additionally, we show that the flexible design of our method enables neural simulation from a CAD geometry alone, thereby eliminating the need for costly CFD meshing procedures for inference.

Figures (13)

• Figure 1: AB-UPT is a neural surrogate model trained to jointly model surface and volume variables of automotive CFD simulations with $>100$M simulation mesh cells. It obtains state-of-the-art surface and volume predictions (left), accurately models drag and lift coefficients (center), all on a single GPU (right).
• Figure 2: We introduce problem scale in the context of CFD simulations and develop a neural surrogate architecture that can handle data from publicly available CFD datasets that uses a high-fidelity CFD approach (HRLES) routinely employed in industry-standard automotive simulations.
• Figure 3: Overview AB-UPT.
• Figure 4: The AB-UPT model is a multi-branch Transformer neural operator using three distinct branches: a surface, geometry, and volume branch. The physics blocks within each branch share parameters and interact through cross-attention. To handle problems at industry-scale, we select a subset of $M$ points from the input point cloud. We call these points anchor tokens, and they serve as the keys and values for attention computation. This approach significantly reduces the complexity of attention, making it possible to scale our model to industry-standard simulation meshes. Query positions define where physical quantities are predicted. For simplicity, we assume anchors and queries are disjoint subsets of all positions, enabling evaluation at anchor locations. We will demonstrate in Section \ref{['sec:experiments:stl']} that anchors can also be placed arbitrarily (e.g., on a regular grid), in which case they may fall in regions where evaluating physical quantities is not physically meaningful (e.g., points inside a car geometry).
• Figure 5: Train and test losses with varying number of training inputs. Contrary to numerical solvers, neural surrogates do not require million-scale meshes to train accurate models. Training on more points beyond a dataset-specific threshold results in overfitting and requires larger datasets (either in terms of data-scale, i.e., number of simulations, or problem-scale, i.e., complexity of the problem) to benefit from the increased modeling capacity when using more points, which is costly. This can be observed in the first and second row, where AhmedML and DrivAerML have the same data scale (400 train simulations), but AhmedML has a smaller problem scale, resulting in overfitting. Note that increasing data- or model-scale is orthogonal to our work. Although runtime increases with a larger number of inputs, the amount of GPU-hours required is still low, and scaling to even larger problem-scales could be easily implemented via sequence parallelism liu2023ringattention, which allows scaling quadratic self-attention to millions of tokens. Note that Transformer/AB-UPT requires less memory due to constant memory complexity of FlashAttention dao2022flashattention, whereas Transolver uses memory-heavy torch.einsum operations to create slices. See Appendix \ref{['app:transolver_num_slices']} for more details on Transolver scaling.