(246b) Automated Machine Learning and Microfluidics-Based Study and Optimization of Zeigler-Natta Catalysts

Development of new polymerization processes and the requisite catalysts is a complex task, often taking years and the combined efforts of multiple people, especially when optimizing for various parameters such as optimal chain length, polydispersity index, and economic factors. This challenge has only increased in recent years as there has been a move toward homogeneous Ziegler-Natta catalysts, a class of molecules which provides favorable control and processing characteristics but are notoriously difficult to optimize and understand. Optimization of a Ziegler-Natta catalyst system relies on the careful selection of several key parameters including the solvent, type and concentration of activator, catalyst concentration, monomer addition, temperature and mixing. While all these factors play into generating the desired polymer, the relationship between these variables is often not clear due to the chemistry involved. Proposed mechanisms for this reaction include up to eleven different equilibrium and non-equilibrium reaction steps, the careful balance of which effects the properties of the final polymer. Traditionally this has been reconciled by performing combinatorial experiments to individually judge the contribution of each factor, but these experiments take time and create large amounts of chemical waste. In our research we aim to apply new laboratory techniques including high throughput microfluidic screening and machine learning to screen, understand and optimize these catalysis systems.

By miniaturizing the system and performing the experiment in a microfluidic system we are able to carefully control the flow properties and temperature gradients and tie a key characteristic of the reaction- the exotherm- to the activity of the catalyst. The critical step in the polymerization mechanism is the propagation of the polymer chain where the pi-bond in an alpha-olefin is converted to a sigma bond, releasing large amounts of energy. By monitoring this progression in a microchannel using IR thermography we are able to get a real time idea as to the concentration of monomer left in the system and thus quantify the activity of the catalyst under various conditions. This is supplemented with measurements by a Dynamic Light Scattering (DLS) instrument to quantify the chain length of the polymers, an important industrial characteristic. Overall this microfluidic platform allows us to capture relevant kinetic information in a highly controlled and repeatable fashion.

The second component which comes into play for the fast screening and optimization of these polymerization catalysts an automated reagent delivery system and control software. The reagent delivery system is constructed in such a way as to allow for a continuously variable composition of reactants- necessary for understanding broad reaction spaces. Information from the experiment is fed back to a machine learning algorithm trained on first-principles data which interprets the activity of the catalyst and updates the training. Based on the updated training the algorithm predicts a new theoretical maximum efficiency range for the catalyst and sends that particular composition to the pumping sub-system to perform the experiment. Overall this allows for the fast screening of catalysts and for the derivation of optimal operating conditions without the need for numerous manual experiments.

Overall while this methodology is currently being applied to the academic study of Zeigler-Natta catalysts, the overall goal is to enhance industrial adoption of AI and ML for the development of chemical processes. By being able to integrate highly multidimensional data and derive optimized conditions automatically we hope to enable a new path towards catalyst discovery.