Measuring and Improving Transfer of Neural Surrogates for Vehicle Aerodynamics

Favour Nerrise, Julian Busch, Gurpreet Singh Hora, Martin Sipka, Danielle C. Maddix, Georgia Olympia BrikisarXiv 2026 · arXiv

This paper, a collaboration between Siemens and Stanford University, presents a systematic study on how neural surrogate models—AI-driven alternatives to expensive physics simulations—transfer across different vehicle classes in aerodynamics. The authors identify significant failures in standard fine-tuning and introduce GeoReplay, a novel method that uses geometry-weighted source replay and gradient alignment to adapt models to new vehicle designs without losing their original accuracy.

TL;DR

Standard fine-tuning of aerodynamic surrogates often results in "catastrophic forgetting" of the source domain and fails to beat training from scratch in over half of tested cases. The authors propose GeoReplay, which uses geometric similarity and gradient alignment to successfully adapt pretrained models to new vehicle classes using only a handful of simulations, preserving source accuracy while matching target performance.

What Problem Does This Paper Address?

In automotive engineering, evaluating a vehicle's aerodynamic drag and lift is crucial for fuel economy and stability. Traditionally, this requires Computational Fluid Dynamics (CFD) simulations, which are incredibly slow—often taking hours for a single design. Neural surrogates can predict these flow fields in milliseconds, but they have a major flaw: they don't generalize well.

If you train a model on sedans, it won't work "out of the box" for SUVs. Currently, engineers must launch expensive new training campaigns for every new vehicle class. While pretraining has revolutionized fields like NLP and Computer Vision, its effectiveness in the "physics-heavy" world of 3D aerodynamics hasn't been rigorously benchmarked until now.

Methods

The authors introduce GeoReplay, a three-step adaptation strategy designed to overcome "negative transfer" (where pretraining actually hurts performance) and "catastrophic forgetting" (where the model loses its original knowledge).

Figure 1: GeoReplay (1) weights source simulations by Chamfer distance to target training geometries, (2) fine- tunes on target and source batches, and (3) adds a source gradient only if it aligns with the target gradient.
  1. Geometry Weighting: Instead of replaying random source data, GeoReplay uses the Chamfer distance to identify source simulations that are geometrically similar to the new target vehicle. It uses an entropy-regularized coupling to assign weights, ensuring the model focuses on relevant prior knowledge.
  2. Surrogate Fine-Tuning: The model is fine-tuned on a mixture of target data and these weighted source samples.
  3. Target-Aligned Replay: This is the "secret sauce." The algorithm computes the gradients for both the target and source batches. It only applies the source update if its direction agrees with the target gradient (using a cosine gate). If the source update conflicts with what the new data is trying to teach the model, that update is skipped.

The authors also provide a theoretical risk bound that separates geometric coverage (how similar the shapes are) from physics shift (differences in solvers, turbulence models, or flow conditions).

Datasets

The study uses a massive benchmarking suite involving five different surrogate architectures and five independent CFD datasets.

Overview of the source and target datasets, showing variations in geometry, solvers, and flow conditions.
  • Source: DrivAerML (436 road-car simulations).
  • Targets:
    • DrivAerNet++: Production-style road cars.
  • SHIFT-SUV: Quarter-scale SUVs.
  • AhmedML: Simplified "bluff body" reference shapes.
  • WindsorML: Square-back reference bodies.

These datasets differ not just in shape, but in the underlying physics: they use different CFD solvers (OpenFOAM vs. Volcano), turbulence models (RANS vs. DDES vs. WMLES), and flow speeds.

Key Insights and Ideas

  • Fine-tuning is surprisingly fragile: In nearly 60% of the comparisons, a model trained from scratch on just 10 target simulations outperformed a fine-tuned pretrained model.
  • Geometry isn't everything: Even if a target car looks like the source car, differences in how the CFD was calculated (the "physics shift") can lead to negative transfer. Geometric similarity alone cannot guarantee a successful transfer.
  • Catastrophic forgetting is rampant: Standard fine-tuning causes the error on the original source dataset to skyrocket (e.g., from 0.042 to 0.350 in one case).
  • Gradient alignment works: By filtering out source updates that conflict with the target task, GeoReplay prevents the model from being pulled in two different directions, allowing it to learn the new task while remembering the old one.

Key Results

The benchmarking study across 80 different settings (combinations of surrogates, datasets, and data budgets) showed that:

  • GeoReplay lowered the source error of fine-tuning in 77 out of 80 settings.
  • It reduced source error by a median factor of 3.3x compared to standard fine-tuning.
  • Crucially, it achieved these gains while matching the target accuracy of fine-tuning, effectively providing a "free lunch" for model preservation.
Geometric mean of source and target errors as a function of the training budget. GeoReplay (GR) consistently stays below standard fine-tuning (FT) and training from scratch (SC).

Limitations and Future Work

The authors note that their geometric measures currently omit details like mesh density and specific CFD solver settings, which contribute to the "physics shift." Furthermore, while the study is large-scale, it primarily uses DrivAerML as the source.

Future work could explore multi-dataset pretraining (training on all available vehicle classes simultaneously) to see if a more diverse "foundation model" for aerodynamics reduces the need for complex replay mechanisms. Additionally, extending this to non-steady-state or transient aerodynamics remains an open challenge.

Summary updated

Abstract

Neural surrogates for external vehicle aerodynamics are fast alternatives to physics-based computational fluid dynamics (CFD) simulations. Their foundational limitation is a lack of zero-shot generalization across vehicle classes, which results in an expensive, custom training campaign for each new vehicle design. Similar to pretrained models in other fields, a pretrained surrogate should adapt to a new vehicle class using only a few simulations and stay accurate on its pretraining domain. No prior study has systematically tested this transfer behavior. We conduct a large-scale, controlled benchmarking study on five DrivAerML-pretrained surrogates across four independent target datasets. These datasets differ from the source in geometry, CFD solver, turbulence model, and flow conditions. We find that fine-tuning these surrogates outperforms training from scratch in fewer than half of the comparisons and lowers the accuracy of the surrogate on its source dataset. We introduce GeoReplay, a geometry- and gradient-aligned source replay that keeps source simulations resembling the target in training and applies their updates only when they agree with the target gradient. A risk bound separates geometric coverage from differences in the simulated physics, and we prove that every sufficiently small GeoReplay step decreases the target training objective. GeoReplay lowers the source error of fine-tuning in 77 of 80 settings and matches its target accuracy.
BibTeX
@misc{nerrise2026measuring,
  author = {Favour Nerrise and Julian Busch and Gurpreet Singh Hora and Martin Sipka and Danielle C. Maddix and Georgia Olympia Brikis},
  title = {{Measuring and Improving Transfer of Neural Surrogates for Vehicle Aerodynamics}},
  year = {2026},
  eprint = {2611.24843},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.24843},
}
Figure