Active Learning and AI for Computational and Autonomous Experiments
During the course of research, we are often faced with the question of which experimental or computational measurement to perform next. When these measurements are expensive or time-consuming to perform, this question becomes critical. Active learning is a technique in machine learning that provides a framework to systematically answer that question by finding an optimal set of measurements to maximize knowledge gained (given the results already in hand), thus controlling the cost. Machine Learning can then be placed in control of materials synthesis and characterization in a closed-loop system, with active learning providing experiment design on every iteration. Recent examples of autonomous systems include mapping out phase diagrams as well as choosing which computational experiments to perform. In this tutorial, we will introduce the theory of Gaussian processes and active learning using open source python libraries. Here, Gaussian processes are used to learn from prior data and make predictions for the results of future experiments, with associated uncertainty. Active learning then utilizes these predictions to determine the next best experiment to perform. Hands-on exercises will be provided based on actual use cases.
- Mohammad Soltanieh-ha
- Daniel Samarov, NIST
- Boris Kozinsky, Harvard University
- Jonathan Vandermause, Harvard University
- Gilad Kusne