[Published online Journal of Computer Chemistry, Japan Vol.21, 36-38, by J-STAGE]
<Title:> 少数データを効率的に利用したデータ駆動型反応機構探索
<Author(s):> 塩谷 友希, 嶋田 五百里
<Corresponding author E-Mill:> iori(at)shinshu-u.ac.jp
<Abstract:> Chemical reaction neural network (CRNN) is a machine learning model that enables a data-driven search of chemical reaction mechanisms by incorporating reaction kinetics theory into the neural network architecture. Conventionally, 103 order data were required for searching a simple reaction system, but the reduction in the number of data required for CRNN is expected to enable its application to various experimental systems. In this study, we investigated the number of data required for prediction in the CRNN. The result showed that prediction of the reaction is possible with as few as 180 data by avoiding falling into local optimal solutions. We also confirmed that incorporating the concept of material balance into loss function has effect of reducing the computational complexity.
<Keywords:> Reaction mechanism, Kinetics, Chemical reaction neural network, Physics-informed machine learning, Data-driven
<URL:> https://www.jstage.jst.go.jp/article/jccj/21/2/21_2022-0024/_article/-char/ja/
<Title:> 少数データを効率的に利用したデータ駆動型反応機構探索
<Author(s):> 塩谷 友希, 嶋田 五百里
<Corresponding author E-Mill:> iori(at)shinshu-u.ac.jp
<Abstract:> Chemical reaction neural network (CRNN) is a machine learning model that enables a data-driven search of chemical reaction mechanisms by incorporating reaction kinetics theory into the neural network architecture. Conventionally, 103 order data were required for searching a simple reaction system, but the reduction in the number of data required for CRNN is expected to enable its application to various experimental systems. In this study, we investigated the number of data required for prediction in the CRNN. The result showed that prediction of the reaction is possible with as few as 180 data by avoiding falling into local optimal solutions. We also confirmed that incorporating the concept of material balance into loss function has effect of reducing the computational complexity.
<Keywords:> Reaction mechanism, Kinetics, Chemical reaction neural network, Physics-informed machine learning, Data-driven
<URL:> https://www.jstage.jst.go.jp/article/jccj/21/2/21_2022-0024/_article/-char/ja/
