End-to-end machine learning for experimental physics: using simulated data to train a neural network for object...

Area of Science:

• Computational Physics
• Machine Learning
• Materials Science

Background:

• Modern machine learning (ML) models require extensive training datasets for accurate predictions.
• Generating large, high-quality datasets is time-consuming and often impractical for small-scale research.
• Existing methods face challenges in acquiring sufficient real-world experimental data for training ML models.

Purpose of the Study:

• To demonstrate a 'full-stack' computational solution for training convolutional neural networks (CNNs) using simulated data.
• To enable the effective application of ML models to real-world experimental data, even with limited resources.
• To address the time-intensive nature of dataset generation in ML applications.

Main Methods:

• Developed a computational framework for on-the-fly generation of simulated training data.
• Utilized a noise injection process to create simulated data representative of experimental conditions.
• Applied the trained CNN to analyze topological defect annihilation in liquid crystal freely-suspended films.

Main Results:

• The trained CNN achieved accuracy comparable to human expert annotation for defect analysis.
• Demonstrated a four-orders-of-magnitude improvement in time efficiency compared to manual analysis.
• Validated the robustness of the ML approach for analyzing spatial distribution and count of topological defects.

Conclusions:

• The 'full-stack' simulation-based training method is a powerful and efficient approach for ML applications in experimental science.
• This method significantly reduces the time investment required for dataset creation, making ML more accessible.
• The trained network provides a highly efficient and accurate tool for analyzing complex physical phenomena like topological defect dynamics.