Validating and Utilizing Machine Learning Methods to Investigate the Impacts of Synthesis Parameters in Gold Nanoparticle Synthesis

Abstract

The control over the synthesis of gold nanoparticles is crucial to ensure optimal optical and processing properties, but synthesis is complex and interdependent on many variables such as reducing agent, capping agent, and the amount of gold seeds and precursor. Machine learning offers the prospect of giving insight into this multidimensional problem, but the reason for selecting a certain model is often unclear. Here, we apply tree-based machine learning algorithms on the semi-batch, seed-mediated synthesis of gold nanoparticles in the size range of 20–120 nm to analyze the correlation between synthesis parameters, optical spectra, and size. After testing the validity of the machine learning models by nested cross-validation, the Random Forest model is selected as a simple model that can reproduce the outcome of the synthesis well. In a further analysis by SHAP (SHapley Additive exPlanations), chemical relationships that were not explicitly taught to the model but purely derived from the data analysis are revealed.