Machine learning for experimental design of ultrafast electron diffraction.

Area of Science:

• Materials Science
• Physical Chemistry
• Data Science

Background:

• Ultrafast electron diffraction (UED) provides insights into ultrafast material behavior.
• Manual analysis of large UED datasets is time-consuming and limits real-time experimental control.
• Lack of real-time data hinders in situ tuning of experimental parameters and avoidance of sample damage.

Purpose of the Study:

• To develop and demonstrate machine learning methods for real-time analysis of UED data.
• To enable in situ monitoring and control of material dynamics during UED experiments.
• To accelerate the discovery and optimization of material properties using UED.

Main Methods:

• Convolutional Neural Networks (CNNs) were trained on synthetic and experimental diffraction patterns.
• CNNs were used for real-time analysis to resolve dynamical processes and identify material damage.
• Convolutional Variational Autoencoders (CVAEs) were developed to track structural phase transformations in a latent space.

Main Results:

• Real-time analysis of UED data was achieved using CNNs.
• Dynamical processes and material damage were identified in a representative material.
• CVAE models successfully tracked structural phase transformations through UED image time trajectories.
• Experimental parameters were steered in real-time towards desired phase transformations.

Conclusions:

• Machine learning, particularly CNNs and CVAEs, can perform real-time analysis of UED data.
• This approach enables self-correcting diffraction experiments for optimizing large-scale user facilities.
• Real-time data analysis facilitates in situ control over material dynamics and experimental outcomes.