Reinforcement Learning-Guided Long-Timescale Simulation of Hydrogen Transport in Metals.

Area of Science:

• Materials Science
• Computational Chemistry
• Condensed Matter Physics

Background:

• Atomistic simulations of diffusion in complex alloys face timescale limitations, restricting analysis to durations shorter than experimental relevance.
• Understanding atomic processes in materials is crucial for designing new alloys with desired properties.

Purpose of the Study:

• To develop long-timescale simulation methods using reinforcement learning (RL) to overcome the timescale problem in atomistic diffusion simulations.
• To implement and explain two specialized RL algorithms: RL transition kinetics simulator (TKS) and RL low-energy states sampler (LSS).

Main Methods:

• Development and application of reinforcement learning (RL) techniques for long-timescale atomistic simulations.
• Implementation of RL transition kinetics simulator (TKS) for simulating diffusion processes.
• Implementation of RL low-energy states sampler (LSS) for efficient sampling of material configurations.

Main Results:

• Computed hydrogen diffusivity in pure metals and CrCoNi alloy using RL TKS, achieving good agreement with experimental data.
• Observed counter-intuitive cooperative motion between hydrogen and vacancies, a novel finding in diffusion mechanisms.
• Demonstrated that RL LSS significantly accelerates the sampling of low-energy configurations compared to the traditional Metropolis-Hastings algorithm.

Conclusions:

• Reinforcement learning provides a powerful approach to extend the accessible timescales in atomistic simulations of diffusion.
• The developed RL methods, TKS and LSS, are effective for studying diffusion kinetics and exploring material configurations.
• This work opens new avenues for simulating complex atomic processes in materials science and related fields.