Bayesian Active Learning for Scanning Probe Microscopy: From Gaussian Processes to Hypothesis Learning.

Area of Science:

• Materials Science
• Physics
• Computer Science

Background:

• Advancements in machine learning and programmable scanning probe microscopes (SPMs) are driving interest in automated and autonomous microscopy.
• Developing effective automated microscopy necessitates specialized machine learning methods and a clear understanding of the interplay between physics and machine learning.

Purpose of the Study:

• To discuss the principles of Bayesian active learning (BAL) and its application in optimizing automated microscopy, particularly for SPM.
• To present a framework that balances domain expertise with machine learning for defining experimental goals and protocols.

Main Methods:

• Exploration of Gaussian processes (GPs) for data-driven modeling and Bayesian inference for physics-based models.
• Discussion of advanced methods including deep kernel learning, structured Gaussian processes, and hypothesis learning.
• Application of these frameworks to integrate prior data and discover functionalities within spectral data.

Main Results:

• Demonstration of how BAL frameworks can incorporate prior scientific knowledge and experimental goals.
• Illustration of how machine learning algorithms can translate these goals into specific experimental protocols for SPM.
• Highlighting the capability to explore physical laws during experiments and discover functionalities from spectral data.

Conclusions:

• The presented Bayesian active learning framework offers a universal approach for automated microscopy techniques combining imaging and spectroscopy.
• This framework is particularly impactful for methods involving destructive or irreversible measurements, such as SPM, nanoindentation, and electron microscopy.
• It enables efficient data acquisition and physical discovery by effectively leveraging domain knowledge and machine learning.