Gaussian Moments as Physically Inspired Molecular Descriptors for Accurate and Scalable Machine Learning Potentials.

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Accurate potential energy surfaces (PES) are crucial for molecular simulations.
• Existing methods balance accuracy (ab initio) with computational cost (empirical potentials).
• Machine learning offers a promising route to bridge this gap.

Purpose of the Study:

• To develop a machine learning (ML) method for constructing high-dimensional potential energy surfaces.
• To create an efficient and accurate ML model for molecular property prediction.
• To enable broader applications in computational chemistry and materials science.

Main Methods:

• Utilized feed-forward neural networks for PES construction.
• Developed an extendable, invariant local molecular descriptor based on geometric moments.
• Implemented the descriptor efficiently on graphical processing units (GPUs).

Main Results:

• Achieved accuracy comparable to established ML models in representing chemical and configurational spaces.
• Demonstrated the model's ability to handle all atomic species with a single neural network.
• Showcased the computational efficiency of the GPU-implemented descriptor.

Conclusions:

• The proposed ML method provides accurate and efficient potential energy surfaces.
• This approach facilitates molecular geometry optimization, rate constant calculations, and molecular dynamics.
• The method represents a significant advancement for computational modeling in chemistry and materials science.