Photo of Arnaud Vadeboncoeur

Arnaud Vadeboncoeur

I am a postdoctoral research associate in the Computational Statistics and Machine Learning Lab at the University of Cambridge, working with Prof. Mark Girolami. My research is at the interface of statistical AI and computational mechanics for engineering applications. I completed my PhD at Cambridge with Prof. Fehmi Cirak. You can find my thesis here: Variational Inference and Probabilistic Models for Parametric Partial Differential Equations.

Research Vision

My goal is to bridge the trust gap in applying data-driven machine learning to engineering systems and scientific computing. I develop statistical machine learning methods to build dependable models of our built environment. Recently, I have been focusing on developing uncertainty quantification methodologies through learning highly informative priors from collections of physical systems via generative models.

In particular, in Efficient Prior Calibration From Indirect Data, we learn optimal prior parameters of a pushforward model through a distribution matching approach. Furthermore, we concurrently learn a surrogate model in the form of a Neural Operator through bi-level opitmization to accelerate learning and obtain a scalable algorithm. Building on this, Efficient Deconvolution in Populational Inverse Problems extends this methodology to solve the blind deconvolution problem, where we learn the distribution of higly correlated polluting noise. The proposed methodologies are applied to important problems in engineering and science, namely porous medium flow, elastodynamics, and simplified models of atmospheric dynamics.

Links & Contact

Publications

Preprints

* Corresponding author    ♯ Equal contribution