Gaussian approximation potentials: the accuracy of quantum mechanics, without the electrons

Area of Science:

• Materials Science
• Computational Chemistry
• Physics

Background:

• Accurate interatomic potentials are crucial for simulating materials behavior.
• Traditional potentials often rely on fixed functional forms, limiting their ability to model complex systems.
• Quantum mechanical calculations provide high-fidelity data but are computationally expensive for large-scale simulations.

Purpose of the Study:

• To introduce a novel class of data-driven interatomic potential models.
• To develop models with flexible functional forms capable of capturing complex potential energy landscapes.
• To demonstrate the systematic improvable nature of these models with increasing data.

Main Methods:

• Automatic generation of interatomic potential models from quantum mechanical energies and forces.
• Utilizing machine learning techniques to learn potential energy surfaces.
• Applying the developed models to bulk crystal systems.

Main Results:

• The developed models can be automatically generated and are not restricted by a predefined functional form.
• These potentials accurately model complex potential energy landscapes.
• The models show systematic improvement with additional training data.
• Application to bulk crystals and high-temperature property calculations demonstrated significant computational savings (orders of magnitude) compared to direct quantum mechanical methods.

Conclusions:

• Data-driven, flexible interatomic potentials offer a powerful approach for materials modeling.
• These models significantly reduce the computational cost of molecular dynamics simulations.
• The method enables accurate prediction of material properties at conditions previously inaccessible due to computational limitations.