Lost in Space: Design Manifolds Can Accelerate Design and Optimization Iterations Several Fold
MIE Department Seminar
November 15, 2022
11:00 AM - 12:00 PM
Presenter: Mark Fuge, PhD, University of Maryland
Location: ERF 1043
Abstract: When designing complex geometry like the surface of a turbine blade, engineers face a choice. They can use many surface control points (equivalently, design variables) to achieve subtle changes that can lead to potentially important performance improvements — and run the risk of themselves (or their optimizers) getting lost in the (often exponentially) larger design space that results. Or they can play it safe. Use a lower-dimensional, standard design representation that they can tractably explore and optimize — and run the risk of settling with lower-performance designs. In this talk, I advocate for a different path; one that seemingly gets the best of both worlds. I propose learning a Design Manifold — a low-dimensional, non-linear subspace via Generative Models — that captures the key ways in which a design space varies, and how we have used it to accelerate gradient-free optimization time by 10x compared to traditional representations and 2-3x compared to State of the Art techniques, among other benefits.
Specifically, I'll present recent work that my group has done as part of the ARPA-E DIFFERENTIATE program and NSF's CAREER program, including applications of Inverse Design for Aerodynamic and Heat Transfer surfaces, among other examples. I'll also present recent work on understanding Bayesian Optimization (BO) methods, and specifically, a non-intuitive result that shows how diverse initialization strategies like Latin Hypercube Sampling and related techniques can actually harm BO convergence under certain conditions.
Speaker Bio: Mark Fuge is an Associate Professor of Mechanical Engineering at the University of Maryland, College Park, where he is also an affiliate faculty in the Institute for Systems Research and a member of the Maryland Robotics Center and Human-Computer Interaction Lab. His staff and students study fundamental scientific and mathematical questions behind how humans and computers can work together to design better complex engineered systems, from the molecular scale all the way to systems as large as aircraft and ships using tools from Computer Science (such as machine learning, artificial intelligence, and submodular optimization) and Applied Mathematics. He received his Ph.D. from UC Berkeley and has received an NSF CAREER Award, a DARPA Young Faculty Award, and a National Defense Science and Engineering Graduate (NDSEG) Fellowship. He gratefully acknowledges prior and current support from NSF, DARPA, ARPA-E, NIH, ONR, and Lockheed Martin, as well as the tireless efforts of his current and former graduate students and postdocs, upon whose coattails he has been graciously riding since 2015. You can learn more about his research at his lab’s website: http://ideal.umd.edu.
Date posted
Oct 24, 2022
Date updated
Oct 24, 2022