Active Learning for Continual Learning & Sequential Decision-Making

Sequential decision-making plays a central role in active data collection for complex engineering systems, where simulations, experiments, and physical measurements are often expensive, time-consuming, or limited in availability. Instead of collecting data passively or uniformly, active learning seeks to determine what data should be collected next to improve a model, design, or decision most efficiently.

This research thrust focuses on developing active learning algorithms that leverage quantified model uncertainty to guide the selection of informative samples. These samples may come from high-fidelity simulations, physical experiments, operational data, or multi-fidelity data sources with varying levels of cost and accuracy. By identifying queries that are expected to provide the greatest improvement, these methods can reduce data requirements while improving the reliability of data-driven surrogate models and optimization workflows.

Our work applies sequential decision-making to surrogate modeling, simulation-based design, high-throughput experimental design, and system optimization. These methods support engineering applications where efficient exploration is critical, including data fusion across simulation and experimental sources, design optimization under uncertainty, Digital Twins, and autonomous materials discovery.

Looking forward, this thrust aims to connect active learning algorithms with agentic systems, simulation platforms, and experimental infrastructure. This integration will enable closed-loop self-discovery systems that can plan, query, execute, learn, and improve autonomously.