Machine Learning for PAC1D and SESE

Abstract

This document outlines various machine learning approaches that were taken in an effort to surrogate numerical models Python Ablation Code 1-Dimension (PAC1D) and Scalable Effects Simulation Environment (SESE), with the ultimate objective of discovering the most efficient method for approximating SESE with sparse data utilization. The methods explored include; physics-informed neural networks, deep Galerkin method, deep mixed residual methods, operator network, deep operator network, Fourier neural operator, physics-informed Fourier neural operator, and physics-informed kernel neural operator. Many of the methods showed strengths and weaknesses in their performance, with the physics-informed kernel neural operator showing the most potential for approximating SESE’s behavior.