Pair-distribution-function guided optimization of fingerprints for atom-centered neural network potentials.

Area of Science:

• Computational chemistry and materials science.
• Development of advanced simulation techniques.

Background:

• Atom-centered neural network (ANN) potentials are efficient and accurate for molecular simulations involving chemical reactions.
• The performance of ANN potentials heavily relies on the quality of atomic fingerprints used as input.
• Current fingerprint selection methods may limit the accuracy and efficiency of ANN potentials.

Purpose of the Study:

• To propose and validate a novel optimization strategy for atomic fingerprints.
• To enhance the accuracy and performance of ANN potentials for complex systems.
• To demonstrate the applicability of optimized ANN potentials in advanced simulation methods.

Main Methods:

• Developed an optimization strategy for atomic fingerprints in the f*g space.
• Optimized fingerprints to fit pre-selected template functions for interatomic interactions.
• Applied the optimized strategy to develop an ANN potential for a Pd13H2 nanoparticle system.
• Validated the ANN potential using the adaptive kinetic Monte Carlo (aKMC) method.

Main Results:

• The developed ANN potential for Pd13H2 nanoparticles showed significant improvement over standard potentials.
• The optimized fingerprints led to enhanced accuracy and efficiency in simulations.
• The ANN potential demonstrated sufficient smoothness for use with the adaptive kinetic Monte Carlo method.

Conclusions:

• The proposed fingerprint optimization strategy effectively improves ANN potential performance.
• Optimized ANN potentials can accurately model complex systems and are compatible with demanding simulation techniques.
• This method facilitates the development of more robust and versatile ANN potentials for computational simulations.