Output-weighted and relative entropy loss functions for deep learning precursors of extreme events

Samuel H. Rudy and Themistoklis P. Sapsis

Overview

Software used for study of loss functions with specific aim of accuratley predicting and quantifying extreme events.

Requires

Python version 3.5+ Minimal python requirements: numpy, scipy, matplotlib, tensorflow 2, tensorflow-probability, GPy, numba

Dataset generation requires installation of Nek5000 and dependencies. Run time for Kolmogorov flow simulation is approximately 5 hours on 16 cores and square cylinder is approximately 43.5 hours on 128 cores. Mesh generation uses gmsh.

Data Access

Data files are stored using git large file storage. To download, install git-lfs, clone repository, and use "git-lfs pull".

License

MIT