Machine Learning Approaches for Developing Potential Surfaces: Applications to OH^-(H₂O) (n = 1-3) Complexes

Area of Science:

• Computational Chemistry
• Quantum Mechanics
• Machine Learning in Chemistry

Background:

• Accurate potential energy surfaces are crucial for understanding molecular behavior.
• Traditional methods for calculating these surfaces are computationally expensive.
• Machine learning offers a promising avenue for accelerating these calculations.

Purpose of the Study:

• To develop a novel, two-step machine learning approach for high-level *ab initio* potential surface generation.
• To expand the molecular-orbital based machine learning (MOB-ML) model for learning correlation energies at the complete basis set limit.
• To create accurate neural network potentials for systems like hydroxide and hydronium ions interacting with water molecules.

Main Methods:

• Utilized Gaussian process regression within the MOB-ML model to learn correlation energies.
• Employed smaller basis set orbitals (aug-cc-pVDZ) as features for predicting complete basis set limit energies.
• Integrated MOB-ML with neural network potentials, using diffusion Monte Carlo (DMC) sampled geometries and energies for training.
• Developed protocols to optimize the use of DMC-generated structures in the training process.

Main Results:

• Successfully developed and applied MOB-ML combined with neural networks to generate potential surfaces for OH⁻(H₂O) and H₃O⁺(H₂O).
• DMC calculations using the new potentials showed good agreement with previous results for these floppy molecular systems.
• Generated novel potential surfaces for larger systems: OH⁻(H₂O)₂ , OH⁻(H₂O)₃, H₃O⁺(H₂O)₂, and H₃O⁺(H₂O)₃.
• Found similar proton delocalization levels between hydroxide and hydronium ions bound to the same number of water molecules.

Conclusions:

• The combined MOB-ML and neural network approach provides an efficient route to high-accuracy potential surfaces.
• The developed potentials enable reliable simulations of hydrated proton and hydroxide systems.
• The study highlights similarities in proton delocalization for hydroxide and hydronium systems, consistent with experimental spectral observations.