Riemannian denoising model for molecular structure optimization with chemical accuracy.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Molecular structure optimization is crucial for computational chemistry and materials science.
• Conventional methods often struggle with the complex potential energy surfaces of molecules.
• Existing models typically operate in Euclidean space, which may not be optimal for molecular data.

Purpose of the Study:

• To introduce a new framework for molecular structure optimization using a denoising model on a physics-informed Riemannian manifold (R-DM).
• To improve the accuracy and robustness of molecular energy modeling.
• To demonstrate the advantages of using physics-informed coordinates over traditional Euclidean approaches.

Main Methods:

• Developed a denoising model (R-DM) that operates on a physics-informed Riemannian manifold.
• Leveraged a Riemannian metric better aligned with molecular energy changes.
• Incorporated internal coordinates reflecting energetic properties for enhanced modeling.

Main Results:

• Achieved chemical accuracy with an energy error below 1 kcal/mol.
• Demonstrated improved structural and energetic accuracy on QM9, QM7-X, and GEOM datasets.
• Outperformed conventional Euclidean-based denoising models in comparative evaluations.

Conclusions:

• Physics-informed coordinates offer a powerful approach for complex molecular optimization.
• R-DM framework shows significant potential for advancing computational chemistry and materials science.
• This method provides a more robust way to model potential energy surfaces.