D3-ImgNet: A Framework for Molecular Properties Prediction Based on Data-Driven Electron Density Images.

Area of Science:

• Quantum Chemistry
• Materials Science
• Artificial Intelligence

Background:

• Machine learning is increasingly used for predicting molecular properties.
• Existing models often lack integration with physical mechanisms.

Purpose of the Study:

• Propose D3-ImgNet, a novel deep learning framework for molecular property prediction.
• Integrate data-driven electron density images with physical principles.

Main Methods:

• Developed D3-ImgNet framework combining group theory, DFT mechanisms, deep learning, and multiobjective optimization.
• Utilized QM9 and QM9X datasets for property prediction (atomization energies, dipole moments, forces).
• Applied framework to S_N2 reaction dataset for minimum energy path prediction.

Main Results:

• Achieved high accuracy in predicting molecular atomization energies using the QM9 dataset.
• Demonstrated satisfactory predictive capabilities for dipole moments and forces on the QM9X dataset.
• Successfully predicted minimum energy paths for S_N2 reactions, showing framework adaptability.

Conclusions:

• D3-ImgNet framework effectively predicts molecular properties and reaction pathways.
• Visualizations confirm accurate replication of electron density transfer.
• The framework shows potential for accelerating materials discovery and high-throughput screening.