Graph-Based Machine Learning to Predict Particle Dispersion Evolution from Training Image Sequences

We develop a spatiotemporal particle-based graph neural network (st-PGNN), an image-based graph machine learning framework to predict the shape and size evolution of growing particles as part of particle ensembles. Using crystallization of amorphous fluoroelastomer layers as a model system from atomic force microscopy image sequences. The new method involves implementing a spatiotemporal particle graph neural network. This will enable us to study the impact of neighboring crystallites on the growth kinetics and to obtain quantitative spatiotemporal correlations. The insights from this approach will enable quantification of radial particle growth rates and improve nucleation density statistics. In the long term, our approach is expected to provide data for detailed comparisons to existing particle-growth theories.<br/><br/>Work performed by CO was under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52- 07NA27344.