Robust Conditional Flow Matching for Probabilistic Inverse Problems

A Warren Distinguished Lecture 
with Assad Oberai 

Aerospace and Mechanical Engineering 
University of Southern California

Abstract

Conditional flow matching algorithms have recently been applied successfully to large-scale probabilistic inverse problems. They offer several desirable properties, including the ability to operate with intractable prior and likelihood densities, accommodate non-Gaussian distributions, scale efficiently to high-dimensional settings, and provide an efficient means of computing information gain, which can, in turn, support optimal experimental design. However, these methods also present important challenges. In this talk, Assad Oberai focuses on two such challenges: (i) the tendency to produce collapsed posterior distributions when trained on finite data and (ii) the generation of trajectories that nearly intersect. Oberai proposes potential remedies for each of these challenges. To address posterior variance collapse, Oberai investigates both implicit and explicit regularization strategies for the learned velocity field. To mitigate near-intersections between trajectories, he extends the concept of minibatch optimal transport to the conditional setting. Oberai evaluates the proposed approaches on canonical inverse problems and quantifies the resulting improvements in performance.

Speaker

Assad Oberai is the Hughes Professor and the Chair of Aerospace and Mechanical Engineering in the Viterbi School of Engineering. He earned a Bachelor of Engineering degree from Osmania University (1992), an MS from the University of Colorado (1994), and a PhD from Stanford University (1998) all in Mechanical Engineering. He has held academic appointments at Boston University (2001-2005), Rensselaer Polytechnic Institute (RPI, 2006 -2017) and the University of Southern California (USC, 2018-now) At RPI, he was the interim Associate Dean for Research and Graduate Studies in the School of Engineering, and the Associate Director of the Scientific Computation Research Center. At USC, he was the interim Vice Dean for Research in the Viterbi School of Engineering from 2019 to 2021. 

Oberai leads the Computation and Data Driven Discovery (CD3) group which designs, implements and applies data- and physics-based models and algorithms to solve problems in engineering and science. Problems such as better detection, diagnosis and care of diseases like cancer, understanding the role of physics in medicine and biology, modeling the evolution of multi-physics and multiscale systems like wildfires and hypersonic flows, and reduced-order models for aerospace and mechanical systems. He has authored more than 150 articles in archival journals on these topics, is on the board of academic editors for three journals and is the lead PI for the DoD Multi-University Research Initiative (MURI) titled “Probabilistic Multiscale Inference of Interface Models under Extreme Conditions.”

Oberai is a Fellow of the International Association of Computational Mechanics (IACM, 2026), American Academy of Mechanics (2023), American Society of Mechanical Engineers (ASME, 2020), the American Institute of Medical and Biological Engineering (AIMBE, 2016), and the United States Association of Computational Mechanics (USACM, 2015).  In 2026, he was awarded the Senior Research Award in the Viterbi School of Engineering at USC, and in 2015 he was awarded the Research Excellence Award by the School of Engineering at RPI. He received the Humboldt Foundation Award for experienced researchers in 2009, and the Erasmus Mundus Master Course Lectureship at Universidad Politécnica de Cataluña, Barcelona in 2010. He was awarded the Thomas J.R. Hughes Young Investigator Award for his contributions to Applied Mechanics by the ASME in 2007. He is a recipient of the National Science Foundation Career award in 2005 and the Department of Energy Early Career award in 2004.

Start date
Friday, Dec. 4, 2026, 10:10 a.m.
Location

210 Civil Engineering Building at the University of Minnesota
or
Online (registration is required; information will be sent to your email)

Share