CAREER: Active Representation Learning for Real-World Adaptive Experimental Design

Artificial intelligence is increasingly used to guide scientific discovery, engineering design, and complex decision-making, where each experiment or trial can be costly and time-consuming. A central challenge is how to efficiently identify the most informative experiments from vast and complex design spaces, especially when observations are limited and uncertainty is high. This project develops a new paradigm for adaptive experimental design that enables learning systems to not only model data but also actively decide what data to acquire. The project's novelties are the integration of data representation and experiment selection into a unified learning framework, where the strategy for choosing experiments is itself learned from data rather than specified by fixed rules. The project's broader significance and importance are in accelerating scientific discovery, improving the efficiency of engineering systems, and enabling intelligent decision-making in settings where data collection is expensive or constrained. Technically, the project formulates adaptive experimental design as a coupled optimization problem that jointly learns representations of experiments and policies for selecting new measurements. It develops learning-based acquisition strategies using tools from representation learning, probabilistic modeling, and sequential decision-making. The approach includes methods for uncertainty-aware modeling in high-dimensional settings, architectures that learn to prioritize informative data points, and algorithms that leverage simulation and historical data to train decision policies. It further incorporates multi-fidelity data sources, indirect feedback, and parallel experimentation into a unified framework, enabling scalable and robust decision-making in complex environments. The resulting system is evaluated in applications such as scientific simulation, cyber-physical system optimization, and data-driven protein design, demonstrating improved efficiency and adaptability. This work advances the foundations of data-driven discovery and enables broader adoption of AI in real-world experimental systems. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria. NSF Award ID: 2543755 | Program: 01002627DB NSF RESEARCH & RELATED ACTIVIT,01002930DB NSF RESEARCH & RELATED ACTIVIT,01003031DB NSF RESEARCH & RELATED ACTIVIT | Principal Investigator: Yuxin Chen | Institution: University of Chicago, CHICAGO, IL | Award Amount: $342,196 View on NSF Award Search: https://www.nsf.gov/awardsearch/show-award/?AWD_ID=2543755 View on Research.gov: https://www.research.gov/awardapi-service/v1/awards/2543755.html