Analysis of epoxy properties using machine learning model: Enhancing predictive performance through outlier detection and selective re-experimentation
발표자
()
초록
내용
This study presents a method to enhance predictive performance of data-driven models by improving dataset quality through outlier detection and selective re-experiments. The dataset includes 701 dynamic mechanical analysis (DMA) measurements, focusing on glass transition temperature and crosslinking density. By identifying and selectively re-experimenting on outliers, the study compared the predictive performance of various algorithms (linear, nonlinear, ensemble models) before and after incorporating the re-experimented data. Results showed significant improvement in predictive performance, indicated by reduced root mean square error (RMSE) and increased coefficient of determination (R²) closer to 1. Although this study focused on DMA properties, it is anticipated that this approach can also predict universal testing machine (UTM) properties, such as lap shear strength, directly related to the adhesive performance of epoxy adhesives.