BEGIN:VCALENDAR
METHOD:PUBLISH
PRODID:-//Apple Computer\, Inc//iCal 1.0//EN
X-WR-CALNAME;VALUE=TEXT:USC
VERSION:2.0
BEGIN:VEVENT
DESCRIPTION:Speaker: Agnimitra Dasgupta, USC
Talk Title: Solution of physics-constrained inverse problems using conditional diffusion models
Abstract: Inverse problems involve deducing the cause from observed effects and are ubiquitous across several science and engineering disciplines. Generally ill-posed, an inverse problem often has multiple solutions. The Bayesian paradigm remains popular for the statistical treatment of inverse problems because it is useful for characterizing the relative plausibility of different solutions. However, Bayesian inference is computationally intractable in most practical scenarios. Some recurring challenges include summarizing available data into informative priors, sampling high-dimensional posteriors, and the need for multiple evaluations of a compute-intensive numerical model, likely black-box and mis-specified, for the forward physics. This talk will introduce conditional score-based diffusion models for solving inverse elasticity problems. A conditional score-based diffusion model uses a neural network to approximate the target posterior distribution’s ‘score function’, defined as the gradient of the logarithm of the density. Subsequently, Langevin dynamics enables the generation of new realizations from the target posterior. Training the diffusion model requires a supervised dataset, and forward model simulations can easily construct it. Therefore, the proposed approach is simulation-based and likelihood-free, and there is no need for gradient computations through the forward physics model. Moreover, the diffusion model is reusable for different measurement instances, unlike conventional MCMC-based inference, which amortizes the cost of inference. This talk will demonstrate the efficacy of conditional score-based diffusion model-driven inference on several physics-constrained inverse problems, primarily concerning inverse elasticity problems, that involve synthetic and real experimental data.
Biography: Agnimitra Dasgupta is a Postdoctoral Research Associate in the Aerospace and Mechanical Engineering Department at the University of Southern California (USC). He obtained his Ph.D. in Civil Engineering from USC and a Master's in Civil Engineering from the Indian Institute of Science. Agnimitra's research interest lies at the intersection of uncertainty quantification and scientific machine learning with applications ranging from the health to infrastructure sectors. Agnimitra received the Provost’s Fellowship from USC between 2017 and 2021.
Host: AME Department
More Info: https://ame.usc.edu/seminars/
Webcast: https://usc.zoom.us/j/96060458816?pwd=8LmoG2q6vBCQubqqWpcizd2F1bxqsH.1
SEQUENCE:5
DTSTART:20240918T153000
LOCATION:SSL 202
DTSTAMP:20240918T153000
SUMMARY:AME Seminar
UID:EC9439B1-FF65-11D6-9973-003065F99D04
DTEND:20240918T163000
END:VEVENT
END:VCALENDAR