An explainable machine learning model to predict anion conductivity for accelerating anion exchange membrane research and development Yin Kan Phua, Tsuyohiko Fujigaya, Koichiro Kato Kyushu University, Japan Japan In search for a green power generator, anion exchange membrane fuel cell (AEMFC) that emits zero greenhouse gases during operation is a good choice, except that its core component – anion exchange membrane (AEM) has problems of low ion conductivity and durability. To date, experiment-centric research and development (R&D) has been done abundantly, but it is resource-intensive, consuming huge amount of cost, labor, and time to achieve a breakthrough. Digital technologies such as machine learning (ML) is a hopeful to improve the efficiency of AEM R&D, if problems such as lack of open-source database and knowhow to represent complex AEM polymer structure, such as homopolymer and copolymer, in ML understandable form can be solved. This study reports the use of a self-built AEM database, containing both homopolymer and copolymer, to train ML model for ion conductivity prediction. Explainability of the prediction logic was evaluated using SHAP values 1 . In-house AEM database containing 1211 anion conductivity data for 272 polymers from 62 papers was built. Homopolymer and copolymer count were 15 and 257, respectively. The polymer structures present in the database were converted into ML readable form using Mordred descriptor 2 , and together with experimental conditions such as conductivity measuring temperature, XGBoost, the ML model used in this study, was trained. Validation R 2 of XGBoost was 0.953, meaning that the model managed to grasp the relationship between the complex polymer structure and ion conductivity rather well. The prediction logic of the model was evaluated using SHAP values, where the importance of each explanatory variables was calculated based on its relative impact towards the final predicted value output. Variables that ranked into the top 5 importance were extracted and analyzed, with founding such as ion conductivity measuring temperature having the highest importance, AMID_N_A originating from descriptor the next, and ATSC1are_A coming in the third (Fig. 1). Measuring temperature is known to have a positive relationship with ion conductivity, and the model has successfully grasped such relationship as shown in figure. AMID_N_A stands for averaged molecular identification for nitrogen atom 2 , while ATSC1are_A and MATS2s_A are descriptors related to the topological state of the chemical structures 2 . All three descriptors carry mathematical and topological meaning, but AMID_N_A focuses on the nitrogen atom, which is the ion conducting moiety in AEM, and it was shown that high AMID_N_A has high impact on ion conductivity. Through further analysis of the chemical meaning behind such results, the proposed approach can potentially become an AEM materials guideline.
Fig. 1. SHAP plot showing the top 5 variables deemed important by XGBoost. References 1. S. M. Lundberg et al., NIPS 2017 , 2017 , 4768.H. Morikawa et al., J Cheminform ., 2018 , 10, 1, 4.
P211E
Made with FlippingBook Learn more on our blog