Bayesian Optimization of High-Intensity, Laser-Driven Particle Acceleration Integrating Experiments and Simulations

Project Overview

There is significant interest in developing laser-driven secondary sources due to their applicability to a wide range of fields including neutron radiography, hadron therapy, nuclear physics, and x-ray generation for imaging. The source characteristics depend on many input parameters, some which control the spatial and spectral properties of the laser while others control the plasma target. Exploring the high-dimensional parameter space and finding the optimal configurations for a particular application are often difficult and time-consuming.

With the rise of high-repetition-rate lasers, there has been interest in automating this search using methods such as Bayesian optimization. Along with experimental data to guide the search, there are often simulations, similar experiments, theory, or knowledge from experts which can potentially contribute information. However, when these information sources are poor approximations of the experiment, they can reduce the efficiency of the optimizer. This study investigated methods for integrating multiple information sources into a Bayesian framework where sources with different levels of fidelity are incorporated into a fused model with improved predictive capabilities. Specific care is taken to study methods which compensate for unreliable information sources as often it is not known a priori the information quality. Several approaches are demonstrated on a two-dimensional benchmark function as well as on multi-fidelity, laser-driven ion acceleration simulations.

Mission Impact

This work addresses the R&D goals in High-Energy-Density Science and High-Performance Data Science. The rise of high-repetition-rate lasers and their significance for High Energy Density Science experiments have been recognized by Lawrence Livermore National Laboratory and DOE. Increased repetition rates allow for refined statistics as well as wider and faster exploration over parameter space, improvements which can advance several programmatic efforts. To realize these systems' full potential, rapid optimization loops will be required. This work addresses this need by employing multiple information sources into the optimization loop. Future work can combine these developed methods with feedback to continue the Laboratory's leadership in high-repetition-rate laser systems. While this study focused on laser-driven proton acceleration, the method can be extended to other experiments including laser-driven neutron sources which have a wide applicability to NNSA mission needs.

Publications, Presentations, and Patents

Matt Hill,"Optimizing High Rep-Rate Radiography with Machine Learning" (PowerPoint Presentation, ELI User Meeting, Dolni Brezany, Czechia, November 2, 2022).

Kelly Swanson, "Application of Machine Learning to Guide High-Repetition-Rate, Laser-Driven Particle Acceleration Experiment" (Poster Presentation, NIF and JLF User Group Meeting, Livermore, California, February 21, 2023).

Kellly Swanson,"Prepulse Effects on Laser-Driven Particle Acceleration in Liquid Crystal Films" (Poster Presentation, LaserNetUS Users' Meeting, College Park, Maryland, June 27, 2023).