Selecting Machine Learning Models to Support the Design of Al/CuO Nanothermites.

Abstract

Novel properties associated with nanothermites have attracted great interest for several applications, including lead-free primers and igniters. However, the prediction of quantitative structure-energetic performance relationships is still challenging. This study investigates machine learning methods as tools to surrogate complex physical models to design novel nanothermites with optimized burning rates chosen for energetic performance. The study focuses on Al/CuO nanolaminates, for which nine supervised regressors commonly used in ML applied to materials science are investigated. For each, an ML model is built using a database containing a set of 2700 Al/CuO nanolaminate systems, specifically generated for this study. We demonstrate the superiority of the multilayer perceptron algorithm to surrogate conventional physical-based models and predict the Al/CuO nanolaminate microstructure-burn rate relationship with good efficiency: the burn rate is estimated with less than 1% error (0.07 m·s-1), which is very good for designing nano-engineered energetic materials, knowing that it typically varies from approximately 8-20 m·s-1. In addition, the optimization of the Al/CuO nanolaminate structure for burn rate maximization through machine learning takes a few milliseconds, against several days to achieve this task using a physical model, and months experimentally.