Trends and Challenges in "Machine Learning" (Workshop 1)

Machine learning is permeating virtually every field of research at an unprecedented pace, and engineering mechanics is of course no exception. Challenges we have been tackling with traditional tools — e.g. accelerating expensive multiscale/multiphysics simulations, solving inverse problems for model calibration, assimilating experimental data, optimizing material design and performance — can now be approached from the perspective of machine learning and its impressive toolbox of models. At the same time, advanced machine learning systems are often portrayed in the news and popular culture as quasi-sentient entities of mystical, mysterious and sometimes insidious nature. Are these models as inscrutable and unpredictable as they sometimes seem? Can we safely let go of some of our trusted physics knowledge in exchange for speed and flexibility?

In this workshop, we take you on a journey across the vast landscape of machine learning techniques used to construct new solutions to engineering mechanics problems. We start in true EM spirit by placing a complicated deep learning model under the microscope, observing its internal structure across the scales, and discussing the blessing and the curse of working with models of virtually unlimited flexibility. We illustrate the classic tradeoff between bias and variance when constructing
machine learning models and explore several strategies to build trustworthy models with an optimum balance between physics and data. Our speakers will showcase a number of exciting research avenues currently being explored by the EM community at the interface between machine learning and engineering mechanics, including experimental and numerical challenges at multiple spatial and time scales.