Chemistry-Informed Generative Model for Classical Dynamics Simulations.

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry
• Generative Models for Molecular Systems

Background:

• Traditional methods for calculating molecular properties, such as *ab initio* calculations and classical dynamics simulations, are computationally expensive.
• Developing efficient methods to generate accurate molecular data is crucial for advancing chemical research.
• Generative models offer a promising avenue for accelerating these calculations.

Purpose of the Study:

• To propose and evaluate a chemistry-informed generative adversarial network (CI-GAN) for generating molecular data.
• To develop an image-input algorithm for simplifying the creation of input databases for complex molecular systems.
• To assess the CI-GAN's ability to predict classical dynamics and *ab initio* calculations.

Main Methods:

• Development of a chemistry-informed generative adversarial network (CI-GAN).
• Implementation of an image-input algorithm for direct molecular image recognition.
• Testing and analysis on representative chemical systems: H + H₂, OH + HO₂, and H₂O/TiO₂(110).

Main Results:

• The CI-GAN approach successfully generates distributions of geometry and energy for the tested molecular systems.
• Chemistry constraints enable the CI-GAN to produce 50%-80% meaningful results from the generated data.
• The model demonstrates potential for predicting classical dynamics and *ab initio* calculations, significantly reducing computational cost.

Conclusions:

• The proposed CI-GAN is a powerful tool for generating accurate *ab initio* energies and molecular dynamics trajectories.
• The image-input algorithm simplifies data preparation for complex molecular systems.
• CI-GANs show significant potential to accelerate chemical simulations and reduce computational expenses.