Data-driven modeling for physical property prediction of polypropylene composites using artificial neural network and principal component analysis

Recent research with artificial neural network (ANN)-based predictive model has emerged as a solution to reduce the number of trial and error effectively. However, it is still challenging to develop a high-performance model using a sparse dataset. Especially, the high-dimension and the small number of polypropylene composite's (PPC's) material data make it difficult to develop a predictive model. In this study, we proposed the ANN-based predictive model using principal component analysis (PCA) to predict the physical property of PPC with the high performance. The optimal dimension reduction of the raw dataset was suggested by the proposed framework to overcome incomplete dataset of PPC materials including the zero values. The dimension reduced dataset was used to develop the ANN-based model for physical property prediction of PPC. As a result, the model accuracy based on the reduced dataset is 0.9061, and 4.6% higher than the model using the raw dataset. This result demonstrates that ANN-based model with dimension reduction improves the prediction performance by reducing the sparsity of PPC material data. Moreover, the proposed model is expected to reduce the number of trial and error in the PPC development process.