Data-Driven Engineering Design: Data-assisted high-fidelity modeling for systems design and monitoring

Increased availability of large amounts of data, development of new sensing technologies and connected systems have generated tremendous interest in the development of methods to learn from data. In parallel, engineers have a long history of building high-fidelity models, based on physical principles, that allow us to model the behavior of highly complex systems. This talk aims at presenting some of the exciting research opportunities that arise at the intersection of these two topics, data analytics, and high-fidelity modeling, and illustrate potential applications. 

For instance, system identification methods make use of measurements from a system to learn the equations and parameters characterizing it, possibly in real-time. Such learning methods find a wide variety of applications, such as calibration and validation of physics-based models of systems during their design process, or diagnostics of mechanical systems. Bayesian inference algorithms are particularly attractive in this framework as they allow quantification of uncertainties around the estimated quantities of interest, where large uncertainties can arise when the data is not highly informative, or the inputs or other experimental variables are uncertain. However, Bayesian techniques become highly computationally expensive when used for inference in large dimensional nonlinear systems, for instance, systems represented by finite element models. This talk aims at demonstrating the potential of Bayesian filtering techniques, such as nonlinear Kalman filters or particle filters, with respect to parameter estimation, and present algorithmic enhancements that achieve a good trade-off between accuracy and computational cost. Furthermore, we will show how these probabilistic learning procedures can also be integrated into more complex frameworks, such as model selection and optimal design of experiments.

The combination of data-mining and physics-based modeling also finds applications in various engineering fields. In the materials sciences, for instance, machine learning algorithms can be used in conjunction with experiments or high-fidelity simulations to obtain a better understanding of structure-property-performance linkages, enabling efficient microstructure characterization and inverse design of materials. Very interestingly, the application of machine learning algorithms to engineering problematics gives rise to challenging problematics and technical questions, for example, the need to accurately quantify uncertainties through machine learning pipelines, or handling small amounts of data. This talk thus aims at illustrating some of the challenges and opportunities related to the use of both model-based and data-based learning algorithms for engineering applications. 

Bio of speaker: Audrey Olivier pursued her undergraduate studies in her native France before moving to Columbia University, New York, for graduate studies. She holds a Diplôme d’Ingénieur (graduate degree) from École Centrale de Nantes, France, and a Master’s of Science and Ph.D. from the Department of Civil Engineering and Engineering Mechanics at Columbia University. Her doctoral research primarily focused on the development of efficient system identification and uncertainty quantification tools, such as Bayesian inference algorithms, for inverse modeling of large dimensional complex systems. Dr. Olivier is currently appointed as a Postdoctoral Fellow in the Hopkins Extreme Materials Institute at Johns Hopkins University. She is working on the development of machine learning algorithms for materials modeling. Dr. Olivier received a Teaching Assistant Excellence Award from her Department at Columbia University in 2015, and a Mindlin Scholar in Civil Engineering and Engineering Mechanics award in 2017 for her work as a doctoral student. Dr. Olivier is also involved in outreach programs aiming at introducing young people to STEM fields.