Prediction of particle agglomeration during nanocolloid drying using machine learning and reduced-order modeling.

[Display omitted] • A method to predict nanoparticle agglomeration using machine learning is presented. • Numerical data and images obtained from KMC simulation serve as training data. • Explanatory variables influencing prediction are determined. • Machine learning combined with ROM enables accurate nanocolloid drying predictions. Our previous studies, which employed Kinetic Monte Carlo (KMC) simulation to model nanoparticle agglomeration during nanocolloid drying, revealed a discernible relationship between agglomeration and the relative Peclet number (Pe*). This study addressed the inverse problem by applying machine learning to predict Pe* for nanoparticle agglomeration. To compensate for the small training dataset used during learning, we combined an artificial neural network with various reduced-order modeling (ROM) techniques. We compared the accuracy of the different ROM techniques based on numerical data and images obtained through KMC simulations, finding that the most accurate Pe* estimations along all Monte Carlo steps were obtained with a clustering decomposition technique leveraging image training data. Our findings facilitate the establishment of relationships between particle agglomeration and specific parameters (e.g., material strength or thermal conductivity) via Pe* , which, in turn, can guide experimental interventions, such as the addition of dispersants, to modulate material properties. [ABSTRACT FROM AUTHOR]