Determination of hyper-parameters in the atomic descriptors for efficient and robust molecular dynamics simulations...

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Atomic descriptors are crucial for machine learning (ML) force prediction in materials science.
• High-dimensional descriptors can lead to overfitting and reduced transferability.
• Efficient and robust ML models require optimized descriptor sets.

Purpose of the Study:

• To develop a method for automatic hyperparameter determination in atomic descriptors.
• To achieve accurate ML force predictions with a reduced number of descriptors.
• To enhance the robustness and transferability of ML models for molecular dynamics.

Main Methods:

• Proposing a novel method to automatically determine hyperparameters for atomic descriptors.
• Focusing on identifying an optimal variance threshold for descriptor components.
• Utilizing both two-body and novel split-type three-body descriptors.

Main Results:

• Demonstrated effectiveness across crystalline, liquid, and amorphous SiO2, SiGe, and Si systems.
• Achieved accurate machine learning forces with a reduced descriptor set.
• Enabled efficient and robust molecular dynamics simulations.

Conclusions:

• The proposed method effectively optimizes atomic descriptors for ML force prediction.
• Reduced descriptor sets lead to more efficient and transferable ML models.
• This approach facilitates robust molecular dynamics simulations in materials science.