Fan Zhou1*, Man Fang1, and Xovee Xu1

1University of Electronic Science and Technology of China, Chengdu, 610054, China

EXTENDED ABSTRACT: New energy technologies such as high-efficiency photovoltaic power generation, large-scale energy storage and electrified transportation power have been rapidly developed. Researchers utilized methods including high-throughput experiments, machine learning and material genetic engineering to speed the development of new energy materials. However, existing formula design methods often face the problems of low efficiency and discovered formula cannot meet the practical demands. This work focuses on the data-driven algorithms for electrolyte multi-parameter mapping, and proposes an accurate mapping model between design parameters (e.g., electrolyte components and preparation process) and performance parameters. To solve the parameter-mapping problem and automate model's iterative optimization this work presents a differentiable ensemble learning to support parameter auto-updating and utilizes attention mechanism to learn electrolyte parameter weights. The proposed model has the following advantages: (1) identify key components and parameters that greatly impact the performance of electrolyte; (2) train a reliable and explainable mapping model; (3) estimate correlations between electrolyte parameters by using metric learning; (4) design proper loss functions for electrolyte data; (5) extract non-linear parameter characteristics via neural networks; (6) explain electrolyte formula recommendation by prior knowledge and influence functions; (7) iteratively optimize model's performance and recommendation through electrolyte tests. The proposed model can effectively predict the performance of electrolyte formula and speed the development and design processes of new electrolyte formulas.

Keywords: electrolyte; multi-parameter mapping; machine learning; artificial intelligence.