FEL-UQ

SLAC LCLS UQ Project

Details:

This repo contains quantile regression and Bayesian NN uncertainty quantification results for the SLAC LCLS FEL pulse energy prediction. The data used for this project consists of SLAC LCLS archive data, where a single sample consists of 76 scalar values as inputs (upstream of the pulse energy detector), and a final scalar output value (photon pulse energy at a gas detector).

This project is active.

The Models:

Quantile regression neural networks: The base model simply uses all of the data to make a models to predict the median measured value, and a 2.5% quantile and 97.5% quantile prediction. Interpolation models (models trained on a subset of data then used to predict on another, time-ordered subset) are trained and evaluated in Interp and Interp2.

Bayesian neural network: all BNN model training is contained in a single notebook. A hyperparameter search was done to determine the architecture - notebook will be added shortly. Please reference the Blitz - Bayesian Layers in Torch Zoo for more information.

Authors:

Primary author (to contact with questions) - Lipi Gupta (lipigupta at uchicago.edu)

SLAC Researchers: Aashwin Mishra and Auralee Edelen

Requirements:

Required packages are listed in the environment.yml file.

It is suggested to use the bash script to make the environment. Simply run the script:

./prepare.sh

conda activate feluq

This repo uses git Large File Storage (LFS) so installation of git-lfs may be needed. For git lfs information.

You can create a unique conda environment for this repo by doing the following (manually):

conda env create -f environment.yml