MGPP-DL: deep learning approach for material graph properties prediction.

Area of Science:

• Materials Science
• Computational Chemistry
• Artificial Intelligence

Background:

• Discovering new semiconductor materials for photovoltaic and optoelectronic applications is computationally expensive using DFT.
• Current deep learning models require 3D crystal structures, limiting their use for hypothetical materials known only by composition.

Purpose of the Study:

• To develop a structure-agnostic deep learning approach for predicting material properties directly from chemical composition.
• To accelerate the screening of novel materials for energy applications.

Main Methods:

• Developed MGPP-DL (Materials Graph Property Prediction via Deep Learning), a structure-agnostic model.
• Represented chemical compositions as graphs with nodes encoding 33 elemental properties.
• Utilized Graph Neural Network (GNN) layers and EfficientNet blocks with attention mechanisms for property prediction.
• Trained the model on 389,000 materials from the Materials Project using DFT-GGA, PBE functional, and scaffold-based data splitting.

Main Results:

• MGPP-DL achieved high accuracy with MAE of 0.0178 eV/atom for formation energy and 0.0619 eV for band gap.
• Demonstrated a 67% improvement over state-of-the-art models with robust generalization (R² = 0.9869 and 0.9182).
• Achieved a 10,000x acceleration compared to DFT, enabling high-throughput screening of millions of hypothetical compositions.

Conclusions:

• MGPP-DL offers a highly accurate and efficient structure-agnostic method for predicting material properties.
• The model significantly accelerates the discovery of novel semiconductor materials for photovoltaic and optoelectronic applications.
• This approach facilitates the exploration of vast chemical spaces for materials design.