Permutation invariant potential energy surfaces for polyatomic reactions using atomistic neural networks.

Area of Science:

• Computational Chemistry
• Materials Science
• Chemical Physics

Background:

• Accurate potential energy surfaces (PES) are crucial for understanding and simulating chemical reactions.
• Atomistic neural networks offer a promising route to constructing complex PES.
• The Behler-Parrinello neural network method is a leading approach for fitting these surfaces.

Purpose of the Study:

• To critically evaluate the applicability and accuracy of the Behler-Parrinello atomistic neural network method.
• To assess its performance in fitting reactive potential energy surfaces for key chemical systems.
• To introduce an efficient Monte Carlo approach for selecting mapping functions.

Main Methods:

• Application of the Behler-Parrinello neural network method to three reaction systems: H + H2, H + H2O, and H + CH4.
• Development and implementation of a pragmatic Monte Carlo method for efficient selection of atom-centered mapping functions.
• Validation of potential energy surface accuracy through fitting errors, direct comparison in dynamically important regions, and quantum scattering calculations.

Main Results:

• The Behler-Parrinello method demonstrates high accuracy in fitting reactive potential energy surfaces for the studied systems.
• The proposed Monte Carlo method provides an efficient way to choose appropriate mapping functions.
• The method's accuracy extends to multidimensional potential energy surfaces, including those with dissociation continua.

Conclusions:

• The Behler-Parrinello atomistic neural network method is both accurate and efficient for representing multidimensional potential energy surfaces.
• This computational approach is suitable for systems involving dissociation continua, broadening its applicability in chemical dynamics.
• The integration with an efficient mapping function selection method enhances its practical utility.