Efficient Configuration Sampling for Hybrid Functional DFT Calculations to Train Machine-Learning Potentials:...

Area of Science:

• Computational Chemistry
• Materials Science
• Chemical Physics

Background:

• Machine-learning potentials (MLPs) are crucial for large-scale molecular dynamics (MD) simulations.
• Current MLPs often rely on computationally inexpensive, lower-accuracy methods.
• Training MLPs with high-accuracy quantum chemical calculations (e.g., hybrid functionals) is computationally prohibitive.

Purpose of the Study:

• To develop an efficient active transfer learning scheme for generating high-accuracy training datasets for MLPs.
• To enable large-scale MD simulations using MLPs trained with hybrid functionals.
• To investigate atmospheric secondary aerosol formation and oxidation reactions.

Main Methods:

• Implemented an active transfer learning strategy to intelligently select configurations for expensive hybrid functional calculations.
• Trained MLPs using the generated dataset for atmospheric chemistry simulations.
• Performed nanosecond-scale MD simulations to study cluster formation and oxidation.

Main Results:

• The developed active transfer learning scheme significantly reduces the cost of generating hybrid functional training data.
• MLPs trained with this method achieve accuracy comparable to direct hybrid functional calculations (errors within meV/atom).
• Stable nanosecond-scale MD simulations were successfully performed, leading to the formation of nanometer-size clusters.

Conclusions:

• The proposed active transfer learning protocol enables efficient and accurate high-level atomistic simulations.
• This approach facilitates the study of complex chemical systems, such as atmospheric aerosol formation.
• Paves the way for a general protocol for high-accuracy, large-scale simulations across various chemical domains.