Deep Learning with ICF Experimental Data

With the recent record-breaking yields from the latest shots at the National Ignition Facility (NIF), it is likely that we have not yet fully optimized the Inertial Confinement Fusion (ICF) experimental design.  Machine learning has already demonstrated capability in ICF design discovery and is gaining traction in the community as a powerful tool. Artificial Neural Networks (ANN) have experienced success across numerous fields of science but often require large amounts of data to train and often overfit with small data sets.  Due to the limited amount of ICF experiments, ANNs frequently train on data generated from numerical simulations. These simulations can be limited in capturing all the physics involved with experiments and may lose out on key features needed for performance prediction. 

Here, we present our findings using multiple deep learning models, trained on experimental data, to predict ICF metrics and perform shot design optimization.  We also discuss the methods we used to reduce the chances of overfitting, including latent space usage from autoencoders,ensembling neural networks, regularization, and custom loss functions. Finally, we compare our results to other machine learning models commonly used with small data sets.

 

This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344.  LLNL-ABS-825158