Scientific Machine Learning for Computational Wave Imaging Problems: from Carbon Zero Emissions to Breast Cancer Detection

"AI for Science" (AI4Science) is emerging as a pivotal area within the machine learning community. In this talk, I will focus on a specific scientific problem: computational wave imaging. Computational wave imaging provides a way to infer otherwise unobservable physical properties of a medium (such as internal density and bulk modulus) from measurements of a wave signal that propagates through the medium. Scientific applications include seismic imaging of the earth, acoustic imaging in materials, and ultrasound tomography in medicine. There are currently two main approaches to solving computational wave imaging problems: those based on physics and those based on machine learning (ML). Among conventional physics-based methods, full waveform inversion (FWI) can provide high-resolution, quantitatively accurate, estimates of medium acoustic properties. However, FWI can be computationally expensive and subject to ill-posedness and “cycle skipping” (a kind of ill-posedness that is particular to wave equations). Recently, ML-based computational methods have been developed to address these issues. Some success has been attained when an abundance of simulations and labels are available. Nevertheless, when applied to a moderately different real-world dataset, ML models usually suffer from weak generalizability. In my talk, I will discuss the details of our recent research effort leveraging both data and underlying physics to address the critical issues of weak generalizability and data scarcity. Particularly, I will go through the advantages and disadvantages of our ML techniques in solving scientific problems of monitoring carbon sequestration using seismic inversion and detecting breast cancers using ultrasound tomography.

Bio
https://smileunc.github.io/