Exploration of Language Models for ICF Design

65^th APS DPP Abstract

Determining particle distributions in phase space is an important problem in ICF and plasma physics. If the phase space distribution of a physical system such as an imploding plasma can be determined, then other physical properties can be readily calculated. A straightforward approach is to start with a known distribution of particles and simulate their evolution. However, for a system of many particles, such computations may be NP hard (a popular computer science term which means Non-deterministic Polynomial-time hard). Neural networks may be used as an alternative method to direct first-principles simulation. A neural network is trained on datasets which capture essential information including experimental data. Once the network is trained, it can make predictions based on new data. Here we explore a relatively new and powerful neural network called language models for ICF experimental design. Language models such as ChatGPT have shown an ability to learn language patterns. By training a model on established information and data, we seek to identify patterns in phase space that amount to a high level of scientific understanding. Specializing a model or an AI on specific research areas may also be more effective than more general Large Language Models such as ChatGPT. The specialized language model will be applied to ICF experimental design. Another purpose of the study is to see if by using the ICF-specific language model, humans can better understand the problems that arise in ICF design.

LANL LA-UR-23-26652