Published: Nov. 20, 2021

Subhayan De Subhayan De
Postdoctoral Associate, Smead Aerospace
Monday, November 29 | 12:00 p.m. | Hybrid - N240 and Zoom Webinar - Register Now

Abstract: Models are built to help represent, understand, and further characterize physical systems. In addition, the ubiquitous presence of uncertainties in material properties, geometry, and loading conditions and their influences on the behavior of these physical systems must also be considered. Recently, neural networks are increasingly used as surrogate models for uncertain physical systems due to their high degree of expressiveness. These networks are trained using data collected from numerical and physical experiments. Once trained, neural networks are computationally inexpensive to evaluate and hence can be used for exercises that require many evaluations of the model, such as uncertainty quantification, to reduce the computational cost. However, in practical engineering settings, often training data from computationally expensive high-fidelity simulations (e.g., using fine grid discretizations) or complex experiments are scarce. On the other hand, a large training dataset from inaccurate lower-fidelity simulations (e.g., using coarser grid discretizations) can be inexpensively obtained. In this talk, the use of neural networks for modeling the behavior of engineering systems under uncertainty and their efficient training using a combination of low- and high-fidelity datasets will be discussed. Another machine learning tool, namely the stochastic gradient descent (SGD) algorithm popular for training neural networks, can be utilized to achieve a robust and reliable design of structures under uncertainty across multiple scales. Two novel variants of the SGD algorithm will be introduced in this talk to reduce the computational cost compared to standard approaches like Monte Carlo for the design of structures.

Bio: Dr. Subhayan De is a postdoctoral associate in Smead Aerospace at CU Boulder. His research focuses on physics-based machine learning and design optimization under uncertainty. Subhayan received his PhD in Civil Engineering from the University of Southern California (USC) in 2018, where he was supported by a Viterbi PhD Fellowship, a Gammel Scholarship, and several NSF grants. At USC, he worked on probabilistic model validation, uncertainty quantification, and structural control design. Subhayan also holds an MS in Electrical Engineering from USC and an MEng in Structural Engineering from the Indian Institute of Science. He received his BEng in Civil Engineering from Jadavpur University, Kolkata.