Active Learning for Moiré Assemblies
This script implements an active learning / Bayesian-style acquisition loop for materials discovery. Its purpose is to iteratively choose which material configurations to sample next so that multiple target properties can be accurately predicted with minimal data. Example Setup: A large design space of candidate material configurations (encoded from composition, stacking, rotation, and angle). Several expensive-to-compute or measure target properties:
The script tries to answer: “Which new material configurations should I evaluate next so that my predictive models improve as fast as possible?”
Practically, it trains neural-network surrogate models on a small subset of data, estimates prediction uncertainty and sensitivity, and then selects new data points to evaluate that are most informative, repeating until all models reach acceptable accuracy.