Neural Network Potential Energy Surfaces for Small Molecules and Reactions

Area of Science:

• Computational Chemistry
• Materials Science
• Quantum Mechanics

Background:

• Accurate interatomic potentials are crucial for molecular dynamics simulations.
• Traditional methods struggle to capture complex many-body interactions efficiently.

Purpose of the Study:

• To review neural network (NN)-based methods for constructing molecular potential energy surfaces (PES).
• To highlight NN applications in classical and quantum dynamics, including reaction dynamics and computational spectroscopy.

Main Methods:

• Direct NN fitting of PES from discrete samples (e.g., ab initio energies).
• NN architectures imposing explicit many-body contributions or multibody representations.
• Dimensionality reduction techniques using NNs for low-dimensional PES representations.
• NNs for constructing sum-of-product form PES for quantum dynamics.

Main Results:

• NNs effectively build accurate PES, including many-body contributions for small molecules.
• NNs facilitate dimensionality reduction and low-dimensional function representations.
• Combinations of NNs with other methods (e.g., permutationally invariant polynomials) yield highly accurate PES for larger systems.

Conclusions:

• NN-based PES construction is a powerful approach for molecular dynamics and spectroscopy.
• NNs offer flexibility in handling complex interactions and improving simulation accuracy.
• Future work may focus on scaling these methods to larger, more reactive systems.