Deep Learning Coordinate-Free Quantum Chemistry.

Area of Science:

• Computational Chemistry
• Quantum Mechanics
• Materials Science

Background:

• Quantum chemical properties are crucial for designing materials, catalysts, drugs, and biological probes.
• Traditional methods like density functional theory are accurate but computationally intensive.
• Current deep learning methods often rely on pre-calculated molecular structures, demanding significant time and resources.

Purpose of the Study:

• To demonstrate accurate quantum chemical computations using deep learning in a coordinate-free domain.
• To investigate the impact of graph-encoding architectures on computational performance.
• To explore the representation of quantum chemical properties by local variable models.

Main Methods:

• Developed and applied deep learning models operating in the coordinate-free domain.
• Utilized molecular graphs (atom connectivity) as input, omitting geometric information.
• Compared the performance of different graph-encoding architectures, including Wave (local variable model) and graph convolutional networks (global variables).

Main Results:

• Accurate quantum chemical property computations were achieved without requiring geometric information.
• The choice of graph-encoding architecture significantly influenced model performance.
• The local variable Wave model accurately calculated quantum chemical properties.
• Wave models outperformed global variable graph convolution models for complex molecules.

Conclusions:

• Deep learning on graph encodings enables efficient and accurate quantum chemical computations in a coordinate-free manner.
• Local variable models, such as Wave, are effective for representing certain quantum chemical properties.
• This approach reduces computational cost and time, facilitating the design of novel molecules and materials.