A nested molecule-independent neural network approach for high-quality potential fits

Area of Science:

• Computational Chemistry
• Quantum Chemistry
• Machine Learning Applications

Background:

• Accurate potential energy surfaces are crucial for understanding molecular behavior and dynamics.
• Traditional methods for fitting potential energy surfaces can be challenging, especially for complex molecules.

Purpose of the Study:

• To evaluate the efficacy of neural networks (NNs) for fitting potential energy surfaces (PES) of small molecules.
• To develop and validate a novel 'nested neural network' technique for improved PES fitting accuracy.

Main Methods:

• A simple neural network approach was applied to the H2O molecule.
• A nested neural network technique was developed for fitting the more complex HOOH and H2CO potential energy surfaces.
• The accuracy of the fitted surfaces was validated by calculating vibrational spectra and comparing low-lying energy levels.

Main Results:

• The simple NN approach achieved a root-mean-square error (RMSE) of 1 cm⁻¹ for the H2O surface.
• The nested NN approach yielded an RMSE of 2 cm⁻¹ for the 6-dimensional HOOH and H2CO surfaces.
• Calculated vibrational spectra showed excellent agreement with exact results, with most low-lying levels within 1 cm⁻¹.

Conclusions:

• Neural networks, particularly the proposed nested approach, are efficient and effective tools for fitting molecular potential energy surfaces.
• The nested NN method offers a robust solution for both simple and complex (e.g., double-well) potentials.
• This technique presents a promising universal approach for potential fitting in computational chemistry.