Training atomic neural networks using fragment-based data generated in virtual reality.

Area of Science:

• Computational Chemistry
• Molecular Modeling
• Artificial Intelligence in Chemistry

Background:

• Accurate molecular energy descriptions are crucial for understanding and engineering molecular structures.
• Traditional methods for mapping potential energy surfaces (PES) can be computationally intensive, especially for high-dimensional systems.

Purpose of the Study:

• To introduce a novel paradigm for deriving energy functions of hyperdimensional molecular systems.
• To leverage virtual reality (VR) for efficient data generation and training of atomic neural networks (ANNs).

Main Methods:

• Generating low-dimensional molecular system data within a virtual reality (VR) environment.
• Utilizing this curated data to train atomic neural networks (ANNs) for predicting potential energy surfaces (PES).
• Focusing on chemical reactions with fewer than eight heavy atoms for initial data generation.

Main Results:

• Successfully trained ANNs to predict energies of higher-dimensional systems (nearly 100 atoms) using a small dataset (15k geometries).
• Achieved mean absolute errors of approximately 2 kcal mol^-1.
• Demonstrated the feasibility of generating an ANN-PES for a large reactive radical with a minimal dataset.

Conclusions:

• Virtual reality (VR) enables intelligent curation of high-quality molecular data.
• This approach significantly accelerates the learning process for developing accurate atomic neural network potential energy surfaces (ANN-PES).
• The methodology shows promise for efficient modeling of complex molecular systems.