Note: Read the details of the project are given below. The code is currently under scrutiny for open-source at LANL.
- It combines a conditional variational autoencoder (CVAE) to project high-dimensional data into a lower-dimensional latent distribution. Subsequently, a LSTM network learns the reverse temporal dynamics within this latent space. During the prediction stage, both networks are integrated into an autoregressive loop to estimate all the upstream fields given limited downstream measurements.
-
VAE's latent space encodes the aleatoric uncertainty of the higher-dimensional input dataset. This uncertainity can be propgated through the temporal learner to predict upstream fields with uncertainty bounds.
-
This repository contains codes accompanying the paper. The dataset accompying the paper is available at Zenodo.
