Physics-informed Reinforcement Learning for Expensive Combinatorial Optimization: A Case Study on Nuclear Fuel Design

Abstract: In this seminar, I will give a brief overview of a case study highlighting the intersection between nuclear fuel design optimization and physics-informed reinforcement learning. You will see how we established a connection between game-playing artificial intelligence and nuclear reactor design. You will see us playing video games and draw analogies to allow the computer to learn design tactics of nuclear engineers to accelerate nuclear assembly and core designs [1]. Variety of algorithms such as deep Q learning and proximal policy optimization were benchmarked against genetic algorithms and simulated annealing in solving a complex and expensive combinatorial problem with heavy constraints. We also developed novel hybrid algorithms to reduce computing time and improve optimization performance through a rule-based hybrid approach between reinforcement learning and evolutionary computing [2]. In a two-way collaboration between Exelon corporation and MIT, these research efforts resulted in a new open-source framework called NEORL for our community to enjoy different optimization algorithms. Take a look here https://neorl.readthedocs.io/en/latest/index.html.

[1] Radaideh, M. I., et al. "Physics-informed reinforcement learning optimization of nuclear assembly design." Nuclear Engineering and Design 372 (2021): 110966.

[2] Radaideh, M. I., Shirvan, K. (2021). Rule-based reinforcement learning methodology to inform evolutionary algorithms for constrained optimization of engineering applications. Knowledge-Based Systems, 217, 106836.

Speaker’s Bio: Majdi I. Radaideh has recently started his postdoctoral appointment at the Oak Ridge National Laboratory in November 2021, conducting research on autonomous control, anomaly detection, and uncertainty quantification to reduce mechanical/electrical system interruptions of the Spallation Neutron Source. Before that he was a postdoctoral associate and then research scientist at MIT. He completed his M.S. and Ph.D. in nuclear engineering from the University of Illinois at Urbana Champaign. Radaideh’s research focuses on the intersection between reactor design, nuclear Multiphysics modeling and simulation, advanced computational methods, and machine learning algorithms to drive the advanced reactor research and improve the sustainability of the current reactor fleet. Radaideh has extensive skills in the development and usage of nuclear codes, programming experience, parallel computing, software engineering, and data science platforms. Radaideh is the leading author of +25 journal articles, +50 research publications, has won +10 awards, and is holding minors in computational science and engineering and applied statistics. In free times, Majdi likes to play basketball, jogging, and watch NBA games; he is a loyal Miami Heat fan!