Gated Linear Model induced U-net for surrogate modeling and uncertainty quantification

We propose a novel deep learning based surrogate model for solving high-dimensional uncertainty quantification and propagation problems. The proposed architecture is developed by integrating the well-known U-net with Gaussian Gated Linear Network (GGLN) referred to as induced or GLU-net. GLU-net treats problem an image regression hence, extremely data efficient. Additionally, it also provides estimates of predictive uncertainty. network less complex 44% fewer parameters than contemporary works. illustrate performance in Darcy flow under sparse scenario. consider stochastic input dimensionality be up 4225. Benchmark results are generated using vanilla Monte Carlo simulation. observe accurate efficient even when no information about structure inputs provided network. Case studies performed varying training sample size robustness approach.