DFTB Simulation of Charged Clusters Using Machine Learning Charge Inference.

Area of Science:

• Computational Materials Science
• Quantum Chemistry
• Machine Learning in Physics

Background:

• Self-consistent charge density functional-based tight binding (SCC-DFTB) is a powerful method for electronic structure calculations.
• Traditional SCC-DFTB requires iterative self-consistent cycles, which can be computationally expensive, limiting the size of systems studied.
• Accurate atomic charges are crucial for many chemical and physical properties, but their precise calculation can be demanding.

Purpose of the Study:

• To introduce a novel modification to SCC-DFTB that bypasses iterative self-consistent charge calculations.
• To develop a machine learning (ML) model for rapid and accurate prediction of atomic charges.
• To enable the investigation of larger atomic ensembles and complex chemical systems with reduced computational overhead.

Main Methods:

• A machine learning algorithm combining a Coulomb model and a neural network was developed to predict atomic charges.
• The ML model takes atomic positions, described by symmetry functions, as input.
• The ML-DFTB approach performs a single diagonalization, approximating the density matrix, energy, and forces.

Main Results:

• The ML-predicted atomic charges closely match exact SCC solutions (within 10^-2 charge units).
• The ML-DFTB method accurately reproduces the potential energy surface (PES) of charged silicon carbide (SiC) clusters.
• Dissociation barriers for ion emission are well-reproduced, indicating the method's suitability for studying charged cluster stability and ion field emission.

Conclusions:

• The ML-DFTB approach offers significant computational savings compared to standard SCC-DFTB without compromising accuracy.
• This method facilitates the study of larger atomic systems, including surfaces and solid-state materials.
• The ML-DFTB approach provides a new avenue for exploring charged cluster dynamics and related phenomena.