Fast Near Ab Initio Potential Energy Surfaces Using Machine Learning.

Area of Science:

• Computational Chemistry
• Quantum Mechanics
• Machine Learning

Background:

• Accurate potential energy surfaces are crucial for studying molecular vibrations.
• Traditional methods for generating these surfaces are computationally expensive.
• Machine learning offers a promising avenue for accelerating these calculations.

Purpose of the Study:

• To develop a cost-effective machine-learning approach for evaluating potential energies in quantum mechanical studies.
• To enable accurate calculations of ground and excited vibrational states for small molecules.
• To create novel potential energy surfaces for systems lacking them, such as C2H5+.

Main Methods:

• Utilized molecular-orbital-based machine learning (MOB-ML) to achieve CCSD(T) accuracy at Hartree-Fock cost.
• Developed GPU-accelerated Neural Network Potential Energy Surfaces (NN-PES) trained on MOB-ML energies from Diffusion Monte Carlo (DMC) simulations.
• Employed the combined NN+(MOB-ML) approach for variational and DMC calculations on water, CH5+, and C2H5+.
• Trained MOB-ML models using geometries from ab initio molecular dynamics (AIMD) trajectories at various temperatures.

Main Results:

• The NN+(MOB-ML) method accurately reproduced vibrational states of water and ground states of CH5+ and its isotopologues.
• A new potential surface for C2H5+ was generated, enabling exploration of its CH stretching vibrations.
• Models trained with higher energy structures showed comparable results for C2H5+.
• The computational cost was significantly reduced without sacrificing accuracy.

Conclusions:

• The NN+(MOB-ML) approach provides a highly efficient and accurate method for constructing molecular potential energy surfaces.
• This technique significantly lowers the barrier for complex quantum mechanical studies of molecular vibrations.
• The developed method opens possibilities for studying novel molecular systems and reaction dynamics.