PREDICTION OF THE DENSITY OF STATES (DOS) OF MATERIALS USING CRYSTAL GRAPH CONVOLUTIONAL NEURAL NETWORKS (CGCNN)

• 主要作者: Fauzi, Akmal
• 格式: Theses
• 語言: Indonesia
• 在線閱讀: https://digilib.itb.ac.id/gdl/view/86938
• 機構: Institut Teknologi Bandung

This study explores the capability of the Crystal Graph Convolutional Networks (CGCNN) model, which was previously used only for predicting single material properties such as Fermi energy, to predict sequential data properties of materials, such as the Density of States (DOS). The method employed is based on a transformer encoder-decoder architecture. This model can be applied to predict the Density of States of materials, including both electronic (eDOS) and phonon (phDOS) DOS. The evaluation results demonstrate that models using GAT/UniMP encoders achieve better performance for predicting phonon density of states (phDOS) in terms of metrics such as R² (0.734), MAE (0.00663), and MSE (0.0001932). However, the CGCNN encoder shows superiority in the Wasserstein Distance (WD) metric, achieving a score of 0.0592, and exhibits faster model training time (49 minutes). Similarly, for electronic density of states (eDOS) predictions, the CGCNN encoder achieves better WD scores (0.208830) and faster training time (1553.2 minutes).Data analysis reveals that metallic materials with fewer elements tend to yield better WD scores, whereas materials containing oxygen elements show lower performance in certain metrics. These findings highlight the capability of CGCNN to handle sequential data, such as material DOS, demonstrating its potential for efficient material design and optimization in the future.