Toward Reaction Vessel Mimicry: Machine Learning-Assisted Automated Exploration of Alkene Polymerization and Its Transferability

Automated reaction path network exploration and product identification through kinetic analysis are essential for mimicking real reaction vessels. A common practice involves using inexpensive semiempirical methods for initial exploration, followed by energy refinement using more accurate density functional theory (DFT) methods. However, semiempirical methods often yield less accurate reaction kinetics, making them unsuitable for efficient exploration and reliable product prediction. Here, we demonstrate the advantages of iterative training of a delta-learning neural network potential (ΔNNP) for automated reaction path exploration. Using ethylene polymerization catalyzed by the ZrCp2CH3+ catalyst as a model system, we achieve DFT-level accuracy by learning the energy difference between DFT and semiempirical methods. Training the ΔNNP on reaction path networks involving one and two ethylene molecules with the catalyst successfully captures all key elementary steps─initiation, propagation, and termination─which can then be extended to study the polymerization of up to six ethylene molecules. Furthermore, a minimally trained ethylene polymerization model provides a robust foundation for propylene polymerization. We also explore the influence of a cocatalyst on the polymerization elementary step network through additional iterative training. Beyond polymerization, this framework can incorporate other ZrCp2-mediated chemistry, such as metallacycle formation, with minimal additional training─yielding several new metallacycle structures. Overall, this iterative training framework is particularly effective for reactions involving repeated analogous elementary steps, such as polymer growth. The approach enables the model to handle increasingly complex reactions, representing an important step toward realistic mimicking of reaction vessels.