Bayesian Learning of Adatom Interactions from Atomically Resolved Imaging Data.

Area of Science:

• Surface science
• Materials science
• Computational physics

Background:

• Atomic structures and adatom geometries on surfaces contain crucial information about formation thermodynamics and kinetics.
• Generative physical models can capture this information, but require accurate parameterization based on experimental data.

Purpose of the Study:

• To develop a workflow for reconstructing atomic and adatom positions from scanning tunneling microscopy (STM) images.
• To optimize parameters of physical models (Ising model) using Bayesian optimization to match experimental observations.
• To utilize the derived generative model for predictions across parameter spaces and validate with experimental data.

Main Methods:

• Machine learning-based analysis of STM images to reconstruct atomic and adatom positions.
• Bayesian optimization to minimize the statistical distance between physical models and experimental observations.
• Optimization of 2- and 3-parameter Ising models for surface ordering.

Main Results:

• Successfully reconstructed atomic and adatom positions from STM images.
• Optimized Ising model parameters and generated a predictive physical model.
• Validated model predictions by comparing simulated morphologies with experimental observations at varying adatom concentrations.

Conclusions:

• The proposed workflow enables the reconstruction of thermodynamic models and their uncertainties from experimental observations of material microstructures.
• This approach provides a powerful tool for understanding and predicting surface phenomena and material behavior.