Data-Driven Polymer Classification Using BiGRU and Hybrid Metaheuristic Optimization Algorithms.

Polymers characterize a different and important class of materials through various industries, all with unique functional properties and structural attributes. Conventional models of polymer classification depend greatly on labor-intensive methods liable to human error and subjectivity. Hence, a continually growing requirement for new polymers with greater properties is a deep understanding and exploration of the chemical space. Hence, data-driven methods for polymers are developing and able to deal with unique challenges originating from the outstanding physical and chemical range of polymers at smaller and larger scales. Recently, Deep Learning (DL) models have considerably transformed material science by allowing for the automatic study and classification of composite polymers. In this paper, a novel optimization algorithm with a DL-Based Neural Networks for Data-Driven Polymer Classification (OADLNN-DDPC) model is proposed. The main intention of the OADLNN-DDPC model is to improve the classification model for data-driven polymers using state-of-the-art optimization algorithms. The data normalization stage is initially executed via Z-score normalization to convert input data into a beneficial format. In addition, the proposed OADLNN-DDPC model implements the bald eagle search (BES) model for feature selection to detect and retain the most appropriate features. For the polymer classification process, the bidirectional gated recurrent unit (BiGRU) technique is employed. Lastly, the zebra optimizer algorithm (ZOA) is implemented for the tuning process. Extensive experiments conducted on a polymers dataset with 19,500 records and 2048 features demonstrated that OADLNN-DDPC achieves an accuracy of 98.58%, outperforming existing models, such as LSTM (83.37%), PLS-DA (88.18%), and K-NN (98.36%). The simulation process of the OADLNN-DDPC model is performed under the polymer classification dataset. The experimental analysis specified that the OADLNN-DDPC model demonstrated improvement over another existing model.