Presentation Information
[2Yin-A-18]Transfer Learning for Detergents Using Self-Supervised LearningFormulation Generation under Constraints and Feature Transfer by Generative Artificial Intelligence
〇Eisuke Inagaki1, Sota Watanabe1, Yu-ich Fujiwara1 (1. Lion Corporation)
Keywords:
LLM,Transfer Learning,SSL,MI,Detergent
The diversification of consumer demands necessitates shortened development periods for detergents, which are essential daily products comprising surfactants, polymers, and solvents with complex mechanisms. Conventionally, high-quality detergent development has required extensive trial and error based on researchers' intuition and experience. We previously reported a transfer learning method using self-supervised learning for virtual screening, achieving high-accuracy quality prediction. However, acquiring large-scale data with high domain similarity for pre-training remains challenging due to limited open databases for practical materials. To address this issue, we propose a data generation method utilizing large language models. By incorporating generated data that explicitly specifies the mechanisms underlying detergent quality manifestation into transfer learning, we demonstrated the possibility of achieving prediction accuracy equivalent to or superior to that obtained with real experimental data.