Discovering new materials, safer chemicals, and better manufacturing processes often depends on running costly experiments in real-world laboratories. In many areas of science and engineering, researchers must choose from a large number of possibilities, where testing each experimental design is expensive in terms of resources consumed. This project seeks to transform the practice of selecting experiments to accelerate engineering design and scientific discoveries. The project will develop novel artificial intelligence (AI) methods to help engineers and scientists to adaptively decide which experiments to run next so that promising discoveries can be found with far fewer trials than traditional trial-and-error methods. The project will also strengthen the future AI workforce by training undergraduate and graduate students through new courses and research opportunities, create open-source tools, and benchmark problems to advance AI and scientific discovery. The overarching goal of this project is to develop a general framework for AI-driven adaptive experimental design over large structured design spaces (e.g., sequences and graphs), where each experiment is expensive and only a small fraction of candidate designs can be tested. The research has four inter-connected thrusts. First, developing new probabilistic surrogate models that map high-dimensional discrete designs to experimental outcomes, so that reliable predictions can be made even from small experimental datasets. Second, designing uncertainty-aware deep learning surrogate models that can use large historical datasets while still producing reliable prediction intervals. Third, developing a unified information-theoretic framework for selecting experiments toward a broad set of scientific goals, including optimizing multiple properties, identifying feasible regions, and finding diverse sets of high-quality candidates. Fourth, developing look-ahead planning methods for physical laboratories that account for