wACSF-Weighted atom-centered symmetry functions as descriptors in machine learning potentials.

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Atom-centered symmetry functions (ACSFs) are crucial for describing molecular geometry in machine learning potentials.
• Conventional ACSFs exhibit undesirable scaling issues with an increasing number of unique elements.
• Accurate descriptors are needed for predicting chemical properties like enthalpies and energies.

Purpose of the Study:

• Introduce weighted atom-centered symmetry functions (wACSFs) as an improved descriptor.
• Compare the performance of wACSFs against conventional ACSFs.
• Investigate automated parameter optimization for descriptors.

Main Methods:

• Developed weighted atom-centered symmetry functions (wACSFs).
• Utilized high-dimensional neural network potentials (HDNNPs) for property prediction.
• Employed the QM9 database containing 133,855 molecules for training and validation.

Main Results:

• wACSFs achieve comparable spatial resolution with fewer parameters than ACSFs.
• wACSFs demonstrate significantly better generalization performance in HDNNPs.
• Automated parameter optimization using genetic algorithms is feasible, though empirical parametrization is sufficient for high accuracy.

Conclusions:

• wACSFs provide a more scalable and efficient descriptor for chemical systems.
• The improved descriptor enhances the accuracy and generalization of machine learning potentials.
• This work facilitates more accurate predictions of chemical properties using machine learning.