Abstract: The complexity of designing, modeling, optimizing, and controlling complex systems in science domains demands a paradigm shift from expert intuition, time-consuming trial-and-error approaches to rigorous, data-driven, automated approaches that can accelerate scientific discoveries. Examples of such complex systems include numerical weather simulation, nuclear reactors, and supercomputers.

In this talk, I will present my team’s recent research efforts on theoretically-grounded and scalable artificial intelligence (AI) and machine learning (ML) algorithms for modeling, optimizing, and controlling complex systems. Specifically, I will present an overview of scalable automated deep learning approaches for surrogate modeling in geophysical and nuclear fusion reactor applications, transfer-learning based ML-approaches for optimizing performance in HPC platforms, and a chance-constrained reinforcement learning method for controlling nuclear reactors with safety constraints. Finally, I will present my perspectives on how large foundation and language models such as GPT-3 can potentially be used to accelerate modeling, optimizing, and controlling tasks in science domains.