Force-Field Optimization by End-to-End Differentiable Atomistic Simulation.

Area of Science:

• Computational materials science
• Atomistic simulations
• Force field development

Background:

• Accurate atomistic simulations rely on precise force fields.
• Traditional parameter optimization methods struggle with complex material properties.
• Current machine learning approaches often focus on energies and forces, limiting property prediction.

Purpose of the Study:

• To develop a framework for optimizing atomistic force fields using differentiable simulations.
• To enable the accurate prediction of complex material properties beyond energies and forces.
• To improve the accuracy and generalizability of force fields for materials simulations.

Main Methods:

• Implemented a framework with inner loop simulations and outer loop optimization.
• Utilized automatic differentiation for analytical gradient computation in property prediction and force field optimization.
• Optimized classical potentials (Stillinger-Weber, EDIP, BKS) and machine-learned potentials.

Main Results:

• Successfully reproduced elastic constants, vibrational density of states, and phonon dispersion for silicon and SiO2.
• Fine-tuned machine-learned potentials to accurately predict radial distribution functions.
• Achieved improved accuracy and generalizability to unseen temperatures compared to traditional methods.
• Demonstrated optimization towards multiple target properties simultaneously.

Conclusions:

• Differentiable simulations offer a powerful tool for advancing materials understanding.
• Analytical gradient computation enhances the efficiency and accuracy of force field optimization.
• The developed framework provides a versatile approach for both theoretical exploration and practical applications in materials science.