Automated Experiment in Materials Synthesis and Characterization

Instructor: Sergei V. Kalinin, University of Tennessee, Knoxville

This 9 day (18 hour) course will be beneficial to those interested in machine learning for experiment automation in materials science. The course is designed for a broad experimental materials science audience, and assumes no prior knowledge of ML/AI methods.

Course description

Recent developments in artificial intelligence (AI) and machine learning (ML) combined with the emergence of Python control interfaces for imaging and characterization tools and laboratory automation and robotics are rapidly creating unprecedented opportunities for the accelerated discovery and optimization of materials. The adoption of the automated experiment workflows in experimental labs from industry to academia is now limited primarily by the knowledge gap required to effectively integrate automated research tools in everyday lab activities. This series of lectures is designed to meet this need. The course will introduce the opportunities opened by the automated tools for experimental materials and ceramics sciences. It will briefly discuss the need for reward-driven workflow planning and provide the overview of instrumental setup, experimental workflow, data management, and modes of human-AI interaction. The core of the course will be dedicated to the Bayesian-Optimization based automated experiment, providing an in-depth physicist's perspective on this framework. Starting from the basic Gaussian Process based Bayesian Optimization, the attendees will be introduced to the structured Gaussian Processes, Deep Kernel Learning, and multifidelity and multitask GP. These will be illustrated for structure-property relationship discovery and learning physical laws via autonomous microscopy, and materials synthesis. Based on these building blocks, we will explore the principle of simple optimization, design of cost-aware workflows, and co-orchestration of multiple tools. In all cases, the emphasis will be made on hands-on tutorials that guide the attendees from the basic concept to the real-world implementations. Finally, the course will introduce the human in the loop workflows and highlight the possible interactions between human and ML/AI agents in the experimental labs.

The tentative outline of the course is as following:

1. Current opportunities for automated experiment in synthesis and characterization: from single tools to cloud laboratories
2. Objectives and rewards for automated experiment
3. Reward driven workflow design, orchestration, and execution
4. Gaussian processes and Bayesian Optimization
5. GP and BO in real world: cost and value
6. Bayesian Inference and structured Gaussian Processes
7. Hypothesis learning
8. Manifold learning, variational autoencoders, and encoders-decoders,
9. Deep Kernel Learning and structure-property discovery
10. DKL, explainable automated experiments, and human in the loop AE
11. Multifidelity structured GP for co-orchestration of multiple tools
12. Future perspectives

Who should attend

This course is intended for researchers in industry and academia interested in machine learning for experiment automation in materials science. The course is designed for a broad experimental materials science audience, and assumes no prior knowledge of ML/AI methods.