Machine Learning Assisted Design for Active Cathode Materials

In the recent years, biomass, such as plant protein, have gain more and more attentions as alternatives to the fossil fuel-based synthetic polymer materials. Soy protein, as a low cost and abundant natural resources, is a promising candidate to be used as a functional material. Soy protein consists of various amino acid residues with different inter/intra molecular interactions, e.g. hydrogen bonds, ionic interactions, hydrophobic interactions and disulfide bonds, etc. The complicated microstructures offer great potentials for soy protein as functional modifier for polymers. Soy protein-based products, such as plastics, adhesives and coatings, have been extensively utilized in industrial and biomedical fields. In our previous studies, we have demonstrated that, by controlling the denaturation conditions and the weight fractions, soy protein could possess different morphologies and properties, therefore, form various interactions with the polymer molecules, and eventually modify their performances. Besides, both experimental and numerical simulation approaches were adopted to analyze the underlying mechanisms of the protein-polymer interactions. In this study, we use denatured soy protein as modifier to alter the crystallinity and properties of Poly(vinylidene fluoride) (PVDF) polymer; the effects of fabrication parameters of the polymer/soy protein blends, including denaturation conditions and weight fractions of protein phase, are evaluated to provide insights for the optimized designs of the functional modifier.

In addition, for the purpose of screening high-performance material formulas, and predicting the quantitative structure-property relationships, a machine learning assisted surrogate model is developed in this work. With the machine learning algorithms, we could efficiently optimize the fabrication parameters of polymer/soy protein blends with improved properties, rather than ‘exhaustive searching’ all the blends from different fabrication parameters through experiments. We herein compare several data-driven machine learning algorithms to design polymer/soy protein blends with optimized dielectric constants/loss. The process includes three steps: learning step, prediction step and reversing step. We firstly train different machine learning algorithms using the experimental dataset. The correlation within fabrication parameters and between crystal structure and dielectric properties can be revealed. Second step is to compare the prediction results from various machine learning models with experimental data to find the best algorithm based on the prediction accuracy and judging criterion R2 (the coefficient of determinant). Lastly, the reversing engineering model is constructed based on the suggestion-test approach: several new blends with the best dielectric properties and their corresponding fabrication parameters are estimated; afterwards, the candidate blends are experimentally fabricated based on the fabrication parameters, and their crystal structures and dielectric properties are measured to validate the machine learning models. The results illustrate that the data-driven machine learning model is promising in finding new polymer/soy protein blends with enhanced dielectric properties.