Highly Efficient Crystallization Studies through Machine Learning and Automation

Crystallization is an important process in a broad range of industries, though studies on this topic remain complicated. Recently, machine learning has been applied to resolve complex issues in chemistry and material science. Here we present a machine learning model integrated into a robotic platform to effectively propose and automatically perform crystallization experiments for organic small molecules. The model was pretrained on around 140000 simulated data generated by a kinetic model, and fine-tuned by over 7000 experimental data obtained on the automated workstation. The improved prediction accuracy and working efficiency of our integrated platform were presented in case studies compared with the traditional approach by humans. A feature contribution analysis demonstrates that this model provides a holistic data-based perspective of promoting and inhibiting influences to crystallization. This work thereby demonstrates the feasibility of applying machine learning techniques to solid-state studies to reduce cost, boost efficiency and deepen understanding.