Prediction for heat deflection temperature of polypropylene composite with Catboost

Recently, polypropylene composites (PPCs) are in the spotlight because of their versatilities in composite industries. Properties of PPCs are determined by numerous physical property values (PPV), among which heat deflection temperature (HDT), polymer's resistance to distortion, is a key indicator. However, enormous trial and error is required to produce PPCs with desired PPV because there is no theoretical equations between material composition and PPV. Hence, to reduce the cost and time of finding material composition to meet the desired PPV, we proposed a machine learning-based PPV prediction model. However, some categorical data which can have an influence on the prediction model performance are included in the dataset, because some of data were from repeated experiments. Therefore, algorithm case study (Multiple linear regression (MLR), XGBoost, and CatBoost) was conducted to develop the optimal HDT prediction model which could process the normal data as well as the categorical data. The performances of each prediction model were evaluated with R² and RMSE. As a result, the CatBoost-based HDT prediction model was proposed as the optimal model to solve the trial and error problem.