Tutorial 10. Active Learning and Decision Making for Automated Experiments

ANCILLARYEVENT · MAR-10T · ID: MAR-10T

Modern physics faces increasingly complex experimental landscapes from high-throughput materials synthesis to real-time microscopy and physical property measurements, where brute-force data collection is no longer practical. Decision-making frameworks and active-learning algorithms turn this challenge into an opportunity by dynamically guiding experiments toward the most informative measurements, rather than exhaustively scanning vast parameter spaces. At the same time, many automated tools from synthesis robots to electron and scanning-probe microscopes now expose open APIs that let ML agents control every aspect of an experiment. By embedding these decision-driven and active-learning approaches into API-enabled workflows, we can create  self-driving  labs that autonomously adjust synthesis recipes, tune beam conditions, or select imaging parameters on the fly. This tutorial will summarize the basic principles of the myopic (Bayesian Optimization) and non-myopic (A*, MCDTs, dynamic programming) decision making, and illustrate applications of these methodologies for materials discovery, automated microscopy, integration of multiple synthesis and characterization tools, and building theory in the loop workflows.

Topics covered: 

• Decision making from Bayesian Optimization to Monte-Carlo Tree Searches
• Gaussian Processes and Bayesian Optimization for Materials Discovery 
• Building autonomous microscopy workflows from structure-property relationship to combinatorial libraries 
• Co-orchestration and co-navigation: integrating multiple tools and building theory in the loop workflows