Hybrid physics-machine learning models for quantitative electron diffraction refinements.

Area of Science:

• Materials Science
• Computational Physics
• Crystallography

Background:

• Accurate electron microscopy simulations are crucial for crystal structure refinement.
• Modeling real-world experimental effects analytically remains a significant challenge.
• Existing methods struggle with scalability and accurately representing experimental variables.

Purpose of the Study:

• To develop a novel hybrid physics-machine learning framework for electron microscopy simulations.
• To enable joint optimization of physical parameters and experimental variables through differentiable simulations.
• To enhance the accuracy and scalability of quantitative crystal structure refinements.

Main Methods:

• Integration of differentiable physical simulations with neural networks.
• Leveraging automatic differentiation for gradient-based optimization.
• Application to three-dimensional electron diffraction (3D-ED) structure refinement.

Main Results:

• Achieved state-of-the-art refinement performance on synthetic and experimental datasets.
• Successfully recovered atomic positions, thermal displacements, and complex thickness profiles with high fidelity.
• Demonstrated superior scalability compared to traditional second-order methods.

Conclusions:

• Differentiable hybrid modeling is a powerful paradigm for quantitative electron microscopy.
• The framework accurately models complex experimental effects, overcoming analytical limitations.
• The modular architecture is extensible to other physical phenomena and microscopy techniques.