Abstract
Surrogate modeling and uncertainty quantification tasks for PDE systems aremost often considered as supervised learning problems where input and outputdata pairs are used for training. The construction of such emulators is bydefinition a small data problem which poses challenges to deep learningapproaches that have been developed to operate in the big data regime. Even incases where such models have been shown to have good predictive capability inhigh dimensions, they fail to address constraints in the data implied by thePDE model. This paper provides a methodology that incorporates the governingequations of the physical model in the loss/likelihood functions. The resultingphysics-constrained, deep learning models are trained without any labeled data(e.g. employing only input data) and provide comparable predictive responseswith data-driven models while obeying the constraints of the problem at hand.This work employs a convolutional encoder-decoder neural network approach aswell as a conditional flow-based generative model for the solution of PDEs,surrogate model construction, and uncertainty quantification tasks. Themethodology is posed as a minimization problem of the reverse Kullback-Leibler(KL) divergence between the model predictive density and the referenceconditional density, where the later is defined as the Boltzmann-Gibbsdistribution at a given inverse temperature with the underlying potentialrelating to the PDE system of interest. The generalization capability of thesemodels to out-of-distribution input is considered. Quantification andinterpretation of the predictive uncertainty is provided for a number ofproblems.