AdaField: Generalizable Surface Pressure Modeling with Physics-Informed Pre-training and Flow-Conditioned Adaptation

Junhong Zou, Wei Qiu, Zhenxu Sun, Xiaomei Zhang, Zhaoxiang Zhang, Xiangyu Zhu · 2026

Read AdaField: Generalizable Surface Pressure Modeling with Physics-Informed Pre-training and Flow-Conditioned Adaptation by Junhong Zou, Wei Qiu, Zhenxu Sun, Xiaomei Zhang, Zhaoxiang Zhang, Xiangyu Zhu (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-07139.txt.

Introduction

Junhong Zou and co-authors published a 2026 arXiv preprint. The preprint covers AdaField. It appeared in the arXiv PAPERS collection for 2026. The surface pressure field of transportation systems, including cars, trains, and aircraft, is critical for aerodynamic analysis and design. In recent years, deep neural networks have emerged as promising and efficient methods for modeling surface pressure field.

Being alternatives to computationally expensive CFD simulations. Currently, large-scale public datasets are available for domains such as automotive aerodynamics. The preprint gives researchers a focused account of the problem and the approach the authors take.

Catalog record

CREATORJunhong Zou, Wei Qiu, Zhenxu Sun, Xiaomei Zhang, Zhaoxiang Zhang, Xiangyu Zhu
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE33 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-07139

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

AdaField: Generalizable Surface Pressure Modeling with Physics-Informed Pre-training and Flow-Conditioned Adaptation Junhong Zou, Wei Qiu, Zhenxu Sun, Xiaomei Zhang, Zhaoxiang Zhang, Xiangyu Zhu Abstract The surface pressure field of transportation systems, including cars, trains, and aircraft, is critical for aerodynamic analysis and design. In recent years, deep neural networks have emerged as promising and efficient methods for modeling surface pressure field, being alternatives to computationally expensive CFD simulations. Currently, large-scale public datasets are available for domains such as automotive aerodynamics. However, in many specialized areas, such as high-speed trains, data scarcity remains a fundamental challenge in aerodynamic modeling, severely limiting the effectiveness of standard neural network approaches. To address this limitation, we propose the Adaptive Field Learning Framework (AdaField), which pre-trains the model on public large-scale datasets to improve generalization in sub-domains with limited data. AdaField comprises two key components. First, we design the Semantic Aggregation Point Transformer (SAPT) as a high-performance backbone that efficiently handles large-scale point clouds for surface pressure prediction. Second, regarding the substantial differences in flow conditions and geometric scales across different aerodynamic subdomains, we propose Flow-Conditioned Adapter (FCA) and Physics-Informed Data Augmentation (PIDA). FCA enables the model to flexibly adapt to different flow conditions with a small set of trainable parameters, while PIDA expands the training data distribution to better cover variations in object scale and velocity. Our experiments show that AdaField achieves SOTA performance on the DrivAerNet++ dataset and can be effectively transferred to train and aircraft scenarios with minimal fine-tuning. These results highlight AdaField’s potential as a generalizable and transferable solution for surface pressure field modeling, supporting efficient aerodynamic design across a wide range of transportation systems. Code — https://github.com/zoujunhong/UniField Figure 1: Overview. AdaField augments the public automotive dataset for model pre-training and then fine-tunes for adapting to other fields with limited data, such as aircraft wings and trains. Introduction In the aerodynamic analysis of transportation systems, including cars, trains, and aircrafts, the surface pressure fields plays a fundamental role in determining key physical quantities such as lift, drag, lateral force, and overturning moment ( Qiu et al. 2025 ) . Accurately capturing high-resolution surface pressure fields is essential for optimizing aerodynamic performance, ensuring structural integrity, and improving fuel efficiency, among other practical concerns ( Gao et al. 2025 ) . Traditionally, surface pressure field simulations are conducted through computational fluid dynamics (CFD), which is expensive, time-consuming, and struggle to scale across the vast design space of shapes and flow conditions encountered in modern transportation systems. Recently, neural networks have emerged as promising and efficient alternatives for surface pressure modeling, benefiting from their capacity to learn complex patterns from large datasets ( Chen et al. 2025 ; Choy et al. 2025 ; Wang et al. 2025 ; Elrefaie et al. 2025 ) . Compared with CFD, which usually requires hundreds of CPU hours to simulate the pressure field, neural networks can typically complete the prediction within one second. However, data scarcity remains a fundamental obstacle to deploying these models broadly in aerodynamic applications. Although extensive public datasets are available for some domains, most notably automotive aerodynamics ( Elrefaie et al. 2025 ; Ashton and Maddix 2024 ; Ashton et al. 2024b ; Ashton et al. 2024a ) , many specialized subfields, such as trains or aircrafts, lack sufficient data. This limitation substantially limits the effectiveness of conventional neural network approaches. Moreover, substantial differences in geometric scale and flow conditions across transportation modalities further complicate model generalization and transfer. For instance, cars typically use driving speed and wind speed as inputs, while aircraft ( Wang et al. 2025 ) often use Mach number (Ma) and angle of attack (AoA), and high-speed trains operate at distinct velocity and geometric scales against cars. As a result, models pre-trained in one domain often perform poorly when applied directly to others, highlighting the urgent need for approaches that enable efficient knowledge transfer and generalization under data scarcity. To address these challenges, we propose the Adaptive Field Learning Framework (AdaField), which focuses on pre-training models on public automotive datasets and adapt them to other data-limited sub-domains, as shown in Figure 1 . AdaField first introduces Semantic Aggregation Point Transformer (SAPT), which serves as a high-performance backbone, integrating local vector self-attention adapted from Point Transformer ( Zhao et al. 2021 ) and a semantic aggregation module adapted from Slot Attention ( Locatello et al. 2020 ) to efficiently extract geometric features. Local vector self-attention restricts attention computation to the k-nearest neighbors, thus reducing memory requirement for processing point clouds with a large number of points. The semantic aggregation module learns to decompose point clouds into parts, capturing structural commonalities across complex surfaces. On the other hand, to improve cross-domain generalization and address the substantial differences in flow conditions and geometric scales among various transportation systems, AdaField introduces two complementary strategies. First, the Flow-Conditioned Adapter (FCA), designed based on adapter ( Houlsby et al. 2019 ) , injects flow-specific information into the model through a lightweight set of trainable parameters, enabling efficient fine-tuning for new aerodynamic regimes. Second, we employ a Physics-Informed Data Augmentation (PIDA) strategy grounded in the Navier–Stokes equations ( Constantin and Foiaş 1988 ) , expanding the training distribution across diverse object scales and velocities. Together, FCA and PIDA enable AdaField to achieve robust transferability and generalization under data-scarce conditions. We validate the effectiveness of AdaField through experiments on DrivAerNet++ ( Elrefaie et al. 2025 ) , a publicly available large-scale CFD dataset for automotive aerodynamics. Our evaluation focuses on the prediction of surface pressure fields, where AdaField consistently achieves SOTA performance across multiple standard metrics. To further demonstrate its cross-domain generalization capability, we evaluate AdaField on two additional datasets representing distinct transportation systems: a train dataset ( Qiu et al. 2025 ) and an aircraft wing dataset ( Wang et al. 2025 ) . In these experiments, AdaField is first pre-trained on DrivAerNet++, and then fine-tuned with only a small number of samples from each target domain. The results show that AdaField exhibits strong few-shot transferability and generalization to novel aerodynamic scenarios, underscoring its practical potential for data-scarce subdomains beyond automotive applications. To summarize, our main contributions are as follows: • We propose AdaField, a framework for predicting the pressure field on the surface of transportation systems. By integrating local vector self-attention with a semantic aggregation module, it can efficiently handle point clouds with a large number of points and maintain high-performance predictions. • We develop FCA to facilitate domain adaptation under substantial differences in flow conditions across transportation systems, as well as PIDA, a data augmentation strategy which expands the diversity of training data in terms of object scale and velocity, enabling generalization to unseen aerodynamic conditions. • We achieve SOTA performance on DrivAerNet++, and also conduct comprehensive cross-domain generalization studies, demonstrating that AdaField, when pre-trained on large-scale automotive data, can be effectively adapted to train and aircraft scenarios with minimal fine-tuning, verifying its practical potential as a solution for aerodynamic design tasks in data-scarce settings. Related Works Surface Pressure Prediction Surface pressure prediction aims to infer the pressure distribution across a vehicle’s surface given its geometry and flow conditions. Traditional CFD software simulates surface fields by numerically solving the governing equations of fluid dynamics, such

Read the complete text