Physically informed artificial neural networks for atomistic modeling of materials

Area of Science:

• Materials Science
• Computational Chemistry
• Physics

Background:

• Atomistic simulations are crucial for materials science, relying on interatomic potentials for energy and force calculations.
• Traditional potentials lack accuracy, while current machine-learning (ML) potentials exhibit poor transferability to novel structures.
• Density Functional Theory (DFT) databases enable accurate ML potential interpolation but not extrapolation.

Purpose of the Study:

• To develop a new class of machine-learning potentials with improved transferability for atomistic simulations.
• To integrate physical principles into machine-learning potentials to overcome limitations of purely data-driven models.
• To create a general-purpose physically informed neural network (PINN) potential for aluminum (Al).

Main Methods:

• Combining a physics-based analytical bond-order potential with neural network regression.
• Developing a physically informed neural network (PINN) potential framework.
• Training and validating the PINN potential on a dataset for aluminum.

Main Results:

• The proposed PINN potential demonstrates significantly improved transferability compared to traditional and purely data-driven ML potentials.
• The approach successfully integrates physical bonding information into the ML model.
• A general-purpose PINN potential for Al was successfully developed, showing high accuracy.

Conclusions:

• Physics-based machine-learning potentials offer a promising direction for advancing atomistic simulations.
• Informing ML potentials with physical principles enhances their reliability and transferability to unexplored material configurations.
• The PINN approach represents a significant step towards more accurate and versatile interatomic potentials.