Michael Buzzy
Advisor: Surya Kalidindi

will defend a doctoral thesis entitled,

Generative Modeling for Inverse Problems in Materials Science

On

Tuesday, April 15th at 1:00 p.m.
CODA Room 230
and/or
Virtually via MS Teams 
https://teams.microsoft.com/l/meetup-join/19%3ameeting_YjdiNTg4OGUtMjli…

Committee
           Prof. Surya Kalidindi – School of  ME, MSE, and CSE (advisor)
           Prof. Aaron Stebner – School of ME and MSE
           Prof. Peng Chen – School of CSE
           Prof. Felix Herrmann – School of CSE, EE
           Prof. Rampi Ramprasad – School of MSE

Abstract
Generative Modeling (also referred to as Generative AI) has unlocked the unprecedented ability to perform approximate Bayesian inference for high dimensional, ill-posed, and non-linear inverse problems. By extracting patterns from data, generative models successfully approximate complex distributions, both conditional and unconditional, allowing for probabilistic techniques to operate in a new, previously intractable, regime. A wide array of scientific pursuits are interested in probabilistic techniques due to their ability to capture uncertainty, regularize ill-posed problems through the specification of a prior, and return families of solutions rather than singular instances. However, not all scientific studies are fortunate enough to be able to support the large data requirements necessary to bring this new paradigm of probabilistic computing to fruition. One such example is materials datasets, especially those that capture the mesoscale structure of materials. These data are incredibly expensive to obtain (and therefore are lacking in quantity), but due to high levels of ill-posedness and uncertainty, stand to benefit greatly from generative modeling approaches. In this dissertation we investigate how we can break this dichotomy through developing new techniques for training generative models in data scarce regimes. These techniques are broadly centered around the idea of generating synthetic datasets which multiply the utility of a very limited amount of ground truth experimental data. We show general approaches for the generation of synthetic datasets which minimize the number of experimental observations required. Novel techniques for generating synthetic spatial datasets for data which naturally decomposes it patterning into local and global components, and finally we show that these synthetic datasets are suitable for calibrating generative models which have real utility for solving long standing inverse problems in materials science.