講演情報
[A-14-26]Retrieval Constrained Multi-Angle Color Prediction and Candidate Screening for Automotive Metallic Basecoats
〇Shun Liu1, Ayako Iwakoshi2, Kentaro Kawata2, Yoshihiro Kawahara1, Mitsuhiro Kamezaki1 (1. The University of Tokyo, 2. Nippon Paint Corporate Solutions Co., Ltd)
キーワード：
Multi-angle color prediction、Formulation design、Automotive metallic basecoat
Automotive metallic basecoats look different at different angles because aluminum flakes and effect pigments reflect light in different directions. We introduce a workflow to predict these color changes and generate candidate formulas. We used 1,388 historical samples with 1,225 coded material-ratio features to predict six-angle CIELAB values. To keep evaluation fair, similar formulas were kept in the same validation group. We evaluated ridge regression, random forests, XGBoost, LightGBM, and CatBoost, and combined their predictions by out-of-fold linear stacking. The ensemble reduced the average five-angle Delta E76 from 19.40 for the best single XGBoost model to 17.46, and achieved 18.01 on a held-out formula group test. For inverse screening, the workflow retrieves samples close to the target color, builds a local material pool, generates non-negative formulas, selects candidates using predicted color error, consistency with nearby ingredients, and formula diversity, and then further optimizes them. The method can generate candidate formulas for expert review before tests, although near-specular and low-angle predictions remain difficult and may require richer process data.