Technical Summary: Drifting Models for Surrogate Flow Modeling

Problem Statement

Computational Fluid Dynamics (CFD) is the standard for optimizing indoor environments and ensuring air quality, offering high-fidelity flow fields. However, its high computational cost prevents rapid design exploration and real-time control. While data-driven surrogate models offer a solution, they face a trade-off: deterministic architectures (e.g., FNOs, U-Nets) are fast but often fail to capture complex, multimodal flow distributions, predicting smoothed expectations instead. Conversely, generative approaches like diffusion models capture physical complexity and uncertainty but suffer from slow inference due to iterative denoising processes requiring hundreds of neural function evaluations (NFE).

Methodology

The authors adapt the generative drifting framework [11] to fluid mechanics to achieve high-quality, single-pass generation. Unlike diffusion or flow matching, which learn iterative noise-to-data trajectories, drifting models evolve a pushforward distribution $q$ during training to match the data distribution $p$ via a drifting field $V_{p,q}$.

Core Adaptations for CFD

To apply drifting to indoor flow prediction, the authors introduced three critical modifications:

1. Latent-Space Drifting via VAE:
   Instead of operating on raw pixel space or using pre-trained image features, the authors trained a convolutional Variational Autoencoder (VAE) on their simulation data. This maps $64 \times 64$ velocity fields to a compact $4 \times 4 \times 16$ latent representation. Drifting occurs within this latent space, preserving structural integrity while avoiding suboptimal pre-trained features.

2. Conditional Architecture:
   The model is conditioned on boundary conditions including inlet/outlet positions, room geometry, and inlet velocity magnitude ($v_{in}$). Two conditioning variants were investigated:

   • Label-based: Treats inlet/outlet as discrete identifiers from a fixed catalogue, embedding them via a learned table. This is restricted to enumerated configurations.
   • Spatial: Replaces identifiers with a three-channel binary mask explicitly encoding inlet, outlet, and obstacle locations. This variant aims to generalize to unseen geometries.
     Both variants embed the scalar velocity via an MLP and combine it with the geometry embedding to drive a Diffusion Transformer (DiT) backbone.

3. Label-Aware Positive Masking:
   To handle data-sparse settings where batches contain mixed conditions, the authors introduced a binary compatibility mask. This gates the attraction kernel in the drifting field, ensuring generated samples only drift toward positive samples with matching (inlet, outlet, geometry) triples. This prevents the model from averaging across incompatible flow regimes.

Baseline Comparison

The drifting models were compared against a Latent Diffusion Model (LDM) baseline [19, 20] operating in the same latent space. The LDM uses a 1,000-step cosine schedule with cross-attention for conditioning and DDIM sampling for inference.

Experimental Setup

• Dataset: 2,025 steady-state 2D CFD simulations of incompressible air flow in a $1.5\text{m} \times 1.5\text{m}$ room, generated using Simcenter STAR-CCM+ (RANS $k$-$\epsilon$ model).
• Configurations: Three geometric setups (empty room, center obstacle, offset obstacle) with systematic variations in inlet/outlet positions (15 each) and inlet velocities (0.1, 0.2, 0.3 m/s).
• Metrics: Evaluated on field accuracy (nRMSE, $R^2$, Cosine Similarity) and flow-structure consistency (Divergence gap, Vorticity gap).

Results

Quantitative Performance

• Accuracy: The diffusion baseline achieved the highest accuracy (nRMSE: 0.059). The label-based drifting model followed closely (nRMSE: 0.068), demonstrating competitive accuracy in a single step. The spatial drifting variant showed lower fidelity (nRMSE: 0.108) and higher variance, particularly in capturing magnitude, though it recovered dominant flow directions.
• Flow Structure: The label-based drifting model maintained competitive flow-structure consistency, with divergence and vorticity gaps comparable to the diffusion baseline. The spatial variant exhibited systematic over-prediction of rotational structures.
• Inference Speed: The drifting model is two orders of magnitude faster than the diffusion baseline. On an NVIDIA A100, the drifting model achieved a latency of 6.74 ms (1 NFE) compared to 1870.2 ms (1000 NFE) for the diffusion model.

Qualitative Observations

• In empty-room scenarios, the label-based drifting model successfully recovered primary flow topologies, including jets and corner recirculation zones, with error distributions comparable to the diffusion baseline.
• In scenarios with obstacles, the spatial-conditioning variant captured essential topology (flow splitting, boundary separation) but struggled with velocity scaling, producing localized over-predictions.

Significance and Claims

The paper posits that conditional drifting serves as a highly efficient alternative to diffusion-based approaches for surrogate flow modeling. The primary significance lies in unlocking real-time CFD surrogates where inference speed is critical, without the iterative computational burden of diffusion.

• Efficiency: The model achieves NFE=1, reducing inference time by ~280x compared to a 1000-step diffusion baseline while maintaining competitive accuracy.
• Generality: While the current label-based implementation is limited to enumerated configurations, the authors propose the spatial-conditioning variant as a promising path toward generalization to unseen geometries, despite current performance gaps.
• Framework: The work establishes a reusable framework for scientific computing surrogates, combining latent-space drifting, spatial-scalar conditioning, and label-aware masking.

The authors conclude that while diffusion remains slightly more accurate, the narrow performance gap combined with the massive speed advantage makes single-step drifting a viable solution for cost-constrained applications. Future work is directed at closing the accuracy gap for spatial conditioning and extending the model to transient 3D regimes.