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

Area of Science:

• Materials Science
• Computer Science

Background:

• Conventional polymer classification methods are labor-intensive and prone to errors.
• There is a growing need for efficient, data-driven approaches to explore the vast chemical space of polymers.
• Deep Learning (DL) models offer powerful tools for automated analysis and classification in material science.

Purpose of the Study:

• To propose a novel Optimization algorithm with a Deep Learning-Based Neural Networks for Data-Driven Polymer Classification (OADLNN-DDPC) model.
• To enhance the accuracy and efficiency of data-driven polymer classification.
• To leverage advanced optimization algorithms for improved polymer characterization.

Main Methods:

• Data normalization using Z-score normalization.
• Feature selection using the Bald Eagle Search (BES) algorithm.
• Polymer classification employing the Bidirectional Gated Recurrent Unit (BiGRU) technique.
• Model tuning utilizing the Zebra Optimizer Algorithm (ZOA).

Main Results:

• The OADLNN-DDPC model achieved a high accuracy of 98.58% on a dataset of 19,500 records and 2048 features.
• Outperformed existing models including LSTM (83.37%), PLS-DA (88.18%), and K-NN (98.36%).
• Demonstrated significant improvement in polymer classification performance compared to other established methods.

Conclusions:

• The proposed OADLNN-DDPC model offers a superior approach for data-driven polymer classification.
• The integration of DL and optimization algorithms effectively addresses challenges in polymer material analysis.
• This data-driven methodology paves the way for more accurate and efficient discovery of novel polymers.