Multi-objective optimization of the epoxy matrix system using machine learning

Abstract

The material properties of the epoxy matrix system are optimized by applying machine learning to a dataset composed of the data on the epoxy composition with the additives and the experimental material properties with missing values before and after curing. First, we construct regression models to predict material properties from the information about the composition of 22 kinds of epoxies, 11 kinds of reactive agents, 8 kinds of alcohols, 15 kinds of curing agents, and 3 kinds of other additives. As machine-learning models, Partial Least Squares Regression, Support Vector Regression, Random Forest Regression, Kernel ridge regression, and Artificial Neural Networks are used. Secondly, desirable compositions are identified by applying the constructed models to many candidates of possible compositions with paying attention to the restrictions of the range of the mixing ratio. Finally, we succeed in making desirable epoxy matrix systems by adopting the identified compositions based on machine-learning predictions, and the usefulness of our approach is clearly shown. Only a few additional experiments allow us to make heat-resistant epoxy matrix systems with high processability and productivity, which have never been achieved for more than 300 experiments with human-choice compositions in a trial-error approach.

Keywords

• Materials informatics
• Epoxy mixture
• Multi-objective optimization
• Inverse analysis