(248k) Gas Phase Reaction Exploration with Reactive Active Learning

Ab initio methods have limitations in studying chemical reactions due to their computational costs. Empirical forcefields are computationally inexpensive but lack bonding terms to model chemical reactions. Bond-order potentials like ReaxFF enable the sampling of bond-breaking and bond-formation within a system. However, these forcefields lack accuracy and require reparameterization for new systems. Machine learning interatomic potentials (MLIPs) can be used to explore numerous reactions at costs comparable to empirical forcefield methods and with near-ab initio accuracy. However, MLIPs are as good at sampling chemical reactions as the data that they are trained on. Data generation schemes that involve performing long MLIP molecular dynamics (MD) simulations are inefficient to enumerate reactive events. We tackle these challenges by formulating an active learning scheme that uses MLIPs and transition-state finding techniques to generate highly informative reactive datasets. This scheme does not require information of reaction pathways or knowledge of final products, thus minimizing human bias. Our MLIPs are trained using the DeePMD formalism with this reactive active learning scheme. We have accurately determined reaction pathways and transition state barriers for gas phase ammonia synthesis and the reaction of methylene imine with water molecules. We found our method to be data efficient as compared to conventional MLIP active learning schemes. Our scheme can determine new minimum energy pathways for various reactions, making it a robust tool for efficiently screening reactions in small molecule design and discovery.