Modified feed-forward neural network structures and combined-function-derivative approximations incorporating...

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry
• Quantum Mechanics

Background:

• Classical molecular simulations assume atom permutations don't alter potential energy.
• Accurate potential energy surfaces (PES) are crucial for molecular dynamics and spectroscopy.
• Developing efficient methods to construct accurate PES is an ongoing challenge.

Purpose of the Study:

• To develop novel feed-forward neural network (NN) architectures that inherently incorporate permutation symmetry.
• To accurately fit the potential energy surfaces (PES) of molecules like H2O and ClOOCl using these symmetric NNs.
• To demonstrate the efficiency of the proposed method in terms of data requirements and accuracy.

Main Methods:

• Implementation of feed-forward neural network structures with built-in permutation symmetry.
• Application of the combined-function-derivative approximation algorithm for accurate PES fitting.
• Training and testing the NN models on H2O and ClOOCl molecular systems.

Main Results:

• Achieved high accuracy in predicting energies and gradients for H2O and ClOOCl.
• Demonstrated significant reduction in the number of required training data points compared to traditional methods.
• Reported low root-mean-squared (RMS) and mean-absolute (MA) energy errors for both test molecules.

Conclusions:

• The developed symmetric neural network models accurately represent molecular potential energy surfaces.
• The approach offers a more data-efficient route to high-fidelity PES construction.
• This method shows great promise for accurate molecular simulations and chemical predictions.