VEHICLE SURFACE PRESSURE PREDICTION FROM
2D SKETCHES VIA A PRE-TRAINED DIFFUSION MODEL

動画引用元論文

VEHICLE SURFACE PRESSURE PREDICTION FROM 2D SKETCHES VIA A PRE-TRAINED DIFFUSION MODEL

著　者
Shunya Nakamura, Osamu Ito
収　録
The 2nd Workshop on Foundation Models for Science: Real-World Impact and Science-First Design（ICLR 2026 Workshop）
発　行
Mar, 2026
要　旨
In the early stages of automotive design, designers explore shape concepts using 2D sketches, yet existing aerodynamic evaluation methods require 3D geometry representations. We propose an end-to-end approach that directly predicts surface pressure coefficient (Cp) distributions from 2D sketch images via image-to-image translation. By fine-tuning a pre-trained image editing diffusion model (Qwen-Image-Edit-2511) with LoRA on the large-scale automotive aerodynamics dataset DrivAerNet++, we achieve Relative L2 Error (Rel L2) = 0.165 and R2 = 0.955 on the test set. Even with training data reduced to 2.2% (128 samples), the model maintains Rel L2 = 0.218, demonstrating applicability in practical settings where CFD simulation costs are prohibitive. We confirm good generalization to unseen base vehicle categories while revealing that prediction quality degrades for shape features absent from training data, highlighting that coverage of shape features matters more than category labels. We also investigate calibrated uncertainty (UQ) via diffusion ensemble, revealing when variance reliably indicates prediction error. These results demonstrate the feasibility of aerodynamic evaluation without 3D geometry representations, opening a path toward aerodynamic feedback at the earliest design stage.

Speaker

Shunya Nakamura
中村舜也 / Shunya Nakamura

その他 執筆論文（共著含む）

生成AIを活用した初心者向け組立作業習熟支援
Supporting Assembly Skill Acquisition for Beginners Using Generative AI

著　者
中村舜也　Shunya Nakamura、平野由彦Yoshihiko Hirano、三坂啓介Keisuke Misaka、 中村拓真　Takuma Nakamura、西村英明 Hideaki nishimura、伊藤修 Osamu Ito
収　録
人工知能学会全国大会論文集 第40回人工知能学会全国大会（JSAI 2026）
Proceedings of the Annual Conference of the Japanese Society for Artificial Intelligence, Vol.40 (2026)
doi:10.11517/pjsai.JSAI2026.0_1I3GS10f04
発　行
Jun,2026
要　旨
製造現場では少子高齢化に伴い暗黙知の継承が困難になり，習熟方法の確立が課題である。本研究では，モーションキャプチャで取得した動作データをDDTWで時間整列して差分を抽出し，速度やジャーク等の指標で習熟度を定量化する。さらに，数値解析結果・作業標準書・姿勢画像を3層構造のコンテキストとして構成しLLMに入力することで，機械学習モデルの事前学習なしに改善すべき部位とタイミングを自然言語で提示する。実証実験では習熟期間の短縮傾向と指導内容の標準化への寄与が示唆された。