ChE: Physics-informed bayesian optimization: A sequential learning framework

Warning Icon This event is in the past.

November 29, 2023
11:30 a.m. to 12:30 p.m.
Engineering, College of 1520- Ford Activities Room
5050 Anthony Wayne
Detroit, MI 48202
Event category: Seminar

ChE seminar:

Physics-informed bayesian optimization: A sequential learning framework for accelerating scientific design and discovery


Dr. Joel A. Paulson, Assistant Professor, Chemical and Biomolecular Engineering Department, Ohio State University


Bayesian optimization (BO) is a powerful tool for optimizing non-convex black-box (also known as derivative-free) functions that are expensive/time-consuming to evaluate and subject to noise in their observations. It turns out that many important real-world science and engineering problems belong to this class such as optimizing high-fidelity computer simulations for calibration or design purposes, auto-tuning hyperparameters in machine learning algorithms, and efficient material and drug discovery. Although BO has been historically deployed as a purely black-box optimizer, its performance is fundamentally limited by the black-box assumption, resulting in significant losses in performance in practice, especially on problems with high-dimensional, intricately constrained design spaces (e.g., molecular property optimization) and datasets that are very sparse and noisy (e.g., data obtained by molecular dynamic simulations). However, in most real-world applications, only a portion of the model is unknown, suggesting that we might be able to break through these observed performance barriers by “peeking inside the box”. For example, when calibrating a physics-based simulator to experimental data, one could (i) leverage the ability to train/test on a strict subset of the available data; (ii) explicitly account for the fact that the loss function must be non-negative and is a sum over many individual terms; and (iii) reuse precomputed values that remain the same in future simulations. In this talk, we describe some new “physics-informed” Bayesian optimization (PIBO) methods that selectively exploit problem structure to achieve state-of-the-art performance. We then show how these methods can be applied to relevant applications including the efficient discovery of high-performance sustainable energy storage materials and fast calibration of genome-scale bioreactor models.


Joel Paulson is the H.C. Slider Assistant Professor of Chemical and Biomolecular Engineering at The Ohio State University where he is also a core faculty member of the Sustainability Institute (SI) and an affiliate of the Translational Data Analytics Institute (TDAI). He received his B.S. in Chemical Engineering (with highest honors) from the University of Texas at Austin in 2011, and his M.S. and Ph.D. degrees in Chemical Engineering from the Massachusetts Institute of Technology in 2013 and 2016, respectively. He was a postdoctoral researcher at the University of California, Berkeley working in systems and control theory from 2016 to 2019. He is the recipient of several awards including the NSF CAREER Award, the AIChE 35 under 35 Award, the Best Application Paper Prize from the 2020 IFAC World Congress, and The Ohio State Lumley Research Award. He is an Associate Editor for Optimal Control Applications and Methods and the Program Coordinator for Area 10d in the CAST Division of AIChE. His current research focuses on development of new theory and methods for data-driven optimization, physics-informed machine learning, and model predictive control, with applications in sustainable energy storage, chemical looping combustion, and non-equilibrium plasma jets.

November 2023