Machine Learning Small Polaron Dynamics.

Area of Science:

• Materials Science
• Computational Chemistry
• Semiconductor Physics

Background:

• Polarons are fundamental to charge transport in semiconductors, influencing material properties and device efficiency.
• Simulating small polaron dynamics requires long timescales, which are challenging for traditional first-principles molecular dynamics due to infrequent hopping events.

Purpose of the Study:

• To develop a computational framework for accurately simulating polaron hopping dynamics at the nanosecond scale.
• To overcome the timescale limitations of conventional methods for studying polaron behavior.

Main Methods:

• Integration of a message-passing neural network with first-principles molecular dynamics under the Born-Oppenheimer approximation.
• Learning the polaronic potential energy surface by encoding the polaronic state.
• Utilizing long-timescale simulations for statistical significance.

Main Results:

• Accurate estimation of polaron mobilities (including anisotropic cases) and activation barriers.
• Successful application to prototypical polaronic oxides, including hole polarons in MgO and electron polarons in TiO2 (pristine and F-doped).
• Results obtained are within experimentally measured ranges.

Conclusions:

• The developed framework enables efficient and accurate simulation of polaron hopping dynamics.
• This approach provides valuable insights into charge transport mechanisms in semiconductors.
• The method has broad applicability for studying polaron-related phenomena in various materials.