Investigation of a Vehicle Aerodynamic Prediction Model Using Latent Representations with Separated Global and Local Geometric Features

Takuji Nakashima･Taichi Okamoto･Jiangdong Miao･Chisato Sasaki･Bisser Raytchev･Takahide Nouzawa (Hiroshima University)･Keigo Shimizu･Yusuke Nakamura･Kohei Seo (Mazda)･Makoto Tsubokura (Kobe University/RIKEN)

In this study, a latent representation of vehicle geometry is constructed by separating the shape into global structure and local features based on spatial resolution and integrating the resulting representations to better retain features associated with local geometric details. Based on this latent representation, an aerodynamic performance prediction model and a neural-operator-based model for high-resolution prediction of the flow field around a vehicle are constructed, and their effectiveness is examined.

Transfer Learning for Efficient Prediction of Automotive Surface Pressure and Stress Using Graph Neural Networks

Tokiya Tanaka (Kobe University)･Junya Onishi (RIKEN Center)･Takuji Nakashima (Hiroshima University)･Makoto Tsubokura (Kobe University)

To improve the efficiency of automotive aerodynamic performance evaluation, GNN-based models that directly predict pressure and stress fields on vehicle surfaces have attracted considerable attention. However, scratch learning for each vehicle type requires costly CFD data generation and model training. In this study, transfer learning using a pretrained model is applied, and its effectiveness is evaluated in terms of training time and prediction accuracy.

Development of a prediction method for leakage noise through door gaps using CAE

Tatsuya Toki･Yuta Ito (Toyota Motor)･Fumihiko Kosaka (Dassault Systemes)･Shiro Yasuoka (Toyota Motor)

Leakage noise, a type of aerodynamic noise, occurs when sound passes through small gaps such as door clearances during high-speed driving and is a key factor in cabin quietness. This study develops a CAE-based prediction method and applies it to a door structure to estimate sound transmission paths and levels into the cabin under acoustic excitation. The predicted results are validated through comparison with experimental measurements conducted under similar conditions, demonstrating the effectiveness of the proposed approach.

Fundamental Study on Inferring Aerodynamic Characteristics by Large Language Models Using Latent Representations of 3D Shapes

Satoru Ito (Nagoya University)･Koji Nishiguchi (RIKEN)･Koichiro Nakaya･Yuuri Ozaki (Nagoya University)･Hiromune Kanamori･Shingo Noritake･Hiroshi Tanaka (Toyota Systems)

To integrate Large Language Models (LLMs) into the automotive aerodynamic design process, it is essential to accurately capture complex, non-linear geometric features. This study proposes an approach that compresses 3D car shapes into latent vectors using DeepSDF and directly inputs them into an LLM to predict drag coefficients (CD). Specifically, we investigate the effect of expanding the dimensionality of the latent vectors on the inference accuracy. Results using the DrivAerNet++ dataset demonstrate that increasing the dimensionality enhances the LLM's physical inference capability, significantly reducing the prediction error without compromising geometric reconstruction accuracy.