Carbon-based molecular properties efficiently predicted by deep learning-based quantum chemical simulation with large...

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Predicting thermodynamic properties of carbon-based molecules is crucial for energy chemistry.
• Traditional methods like density functional theory are computationally expensive.
• Machine learning integration presents opportunities and challenges due to computational costs.

Purpose of the Study:

• To develop an efficient deep learning model for predicting thermodynamic properties of carbon-based molecules.
• To overcome the computational limitations of conventional simulations.
• To advance the integration of data-driven and physics-based modeling.

Main Methods:

• A deep learning-based quantum chemical prediction model was constructed.
• A novel framework encodes 3D molecular information into a large language model (LLM) generating a 2D SMILES string.
• An equivariant learning module was designed for conformational sampling and feature learning.

Main Results:

• The model accurately predicts thermodynamic properties by preserving 3D structural information.
• It offers significantly faster computational speeds compared to conventional simulations.
• The framework enhances structural information for the model beyond 2D topology.

Conclusions:

• The developed model pioneers efficient practical applications in energy chemistry by combining machine learning and quantum chemistry.
• It advances the integration of data-driven and physics-based approaches.
• The study unlocks novel insights into carbon-based molecules through enhanced predictive capabilities.