빅데이터사이언스 연구실

Tires are critical components that directly influence the core dynamic characteristics of a vehicle, and their performance is evaluated with standardized criteria such as wet grip. Wet grip evaluation under real-world conditions is significantly time-consuming, which impedes the development cycle, especially when a large number of candidate tire designs must be examined. Finite Element Analysis(FEA)-based simulation is a common approach to reducing the cost and time of tire development; however, it still suffers from a high computational cost and takes more than 10 hours to evaluate wet grip for a single design candidate. To address this problem, we propose GRSWP, GRaph-based Semi-supervised learning framework for Wet griP prediction, which is designed for a real-world development scenario and addressing the data scarcity problem of wet grip labels. We first construct a tire-graph where nodes represent tires, using their footprint images as features, and edges connect tires with similar design specifications. We then perform unsupervised graph contrastive pretraining to learn robust tire representations from all available data. Finally, supervised fine-tuning is conducted using the limited set of labeled samples to predict wet grip performance. Experimental results with real-world tire development data demonstrate that GRSWP improves prediction accuracy by approximately 77% and dramatically accelerates the prediction process from over 10 hours to 1.5 milliseconds per sample, compared with FEA-based simulation.