DescriptionTalk in the Mathematics in Computation seminar series.
Abstract: Modeling complex physical systems plays a vital role in many science and engineering domains, where computational models inform decisions and guide design, particularly when data is sparse. We develop a Graph-Informed Neural Network (GINN), which forms part of a broader strategy for incorporating domain knowledge into deep learning for complex physical systems. This framework utilizes probabilistic graphical models to embed expert knowledge, available data, and design constraints into a physics-based representation. Next, the hidden nodes of a neural network (i.e., learned features) replace computationally intensive nodes in the probabilistic graphical model. The resulting GINN is a learned statistical surrogate that can cheaply generate a large amount of output data for sensitivity analysis and further uncertainty quantification. Incorporating available domain knowledge into machine-learned models has the potential to reduce data requirements while accelerating training and prediction and enhancing the accuracy, interpretability, and defensibility of the surrogate. As a proof of concept, we build two GINNs of interest in energy storage: (1) a multiscale model of electrical double-layer capacitor dynamics and (2) a nonlinear dynamical system describing biomolecular adsorption. We also discuss mutual information-based global sensitivity analysis methods for interrogating surrogate models to accelerate design cycles.
|Period||20 May 2021|
|Held at||Oak Ridge National Laboratory, United States, Tennessee|
|Degree of Recognition||International|