Flame spray pyrolysis optimization via statistics and machine learning

Abstract

Flame spray pyrolysis (FSP) is an important manufacturing process whereby nanomaterials are produced through the combustion of atomized fuel containing dissolved precursor elements. While FSP has the potential to enable the scalable production of a wide range of next generation energy materials, it also has a multi-scale, multi-physics character, and a large number of processing variables. Optimizing the process for desirable material outcomes by traditional approaches is challenging. In this work, the processing parameter space is explored via a new and efficient methodology that includes statistical methods such as Latin hypercube design of experiments, machine learning surrogate modeling, and Bayesian optimization. As a result, the FSP process is optimized for enhanced performance. Specifically, in-situ particle size measurements are used to tailor the production of silica nanoparticles for a low spread in particle diameters with respect to the mean particle diameter, resulting in an improvement of 25.5% over the baseline within 15 experimental trials. In the process, the analysis reveals distinct domains of primary particle and agglomerated particle formation.