[22a-M206-1~11] 23.1 Joint Session N "Informatics"

[22a-M206-6] Multitask machine learning to predict miscibility and Flory-Huggins χ parameter of polymer solutions

〇Yuta Aoki¹, Teruki Tsurimoto², Stephen Wu^1,3, Yoshihiro Hayashi^1,3, Shunya Minami³, Kazuya Shiratori², Ryo Yoshida^1,3,4 (1.ISM, 2.Mitsubishi Chemical Group, 3.SOKENDAI, 4.NIMS)

Keywords:polymer, informatics, miscibility

We have developed a machine learning model to predict miscibility and Flory-Huggins χ parameter by multitask learning. By randomly splitting the dataset into the training/test sets 10 times, the average AUC higher than 0.93 and the average R² higher than 0.75 are confirmed for binary classification of miscible/in-miscible and for quantitative prediction of χ parameter, respectively. The architecture of our neural network model reflects the generalization of the conventional predicting method with Hansen solubility parameters (HSP). We will also discuss the details of the relationship between HSP and the architecture of our model.