(62g) Reactive Active Learning for Machine Learning Potentials Using Transition-State Finding Methods

Machine learning forcefields possess significant power by accurately mapping molecular structures onto high-dimensional potential energy surfaces, nearing quantum precision, while requiring only a fraction of the computational resources compared to quantum calculations. A notable advantage of most machine learning forcefields is the use of molecular descriptors, which enable encoding of information about chemical reactions without the need to explicitly add bonding terms, unlike most empirical forcefields. However, training machine learning forcefields to accurately account for a wide range of chemical reactions can be extremely inefficient. Therefore, there is a critical need to develop algorithms to optimize the exploration of configurations that are essential to the reactive part of the potential energy surface.

Conventional active learning methods commonly employed in developing forcefields tailored for specific chemical reactions often entail lengthy molecular dynamics simulations to capture reactive events accurately. Alternative methods like enhanced sampling utilizes predefined reaction coordinates to explore diverse configurations. However, this approach introduces human bias in selecting reaction coordinates or collective variables. To address this limitation, we propose a method that automatically generates potential reactions, coupled with transition-state (TS) finding techniques capable of generating training data without specifying reaction pathways or desired products. We employed the DeePMD formalism to train our forcefields in an iterative active learning scheme. These cost-effective forcefields facilitated the exploration of reactive space through automatic generation of potential reactions and employment of TS finding methods such as the single-ended growing string and the nudged elastic band techniques. We evaluated this active learning scheme by testing it on few gas-phase reactions: ammonia synthesis and methylene imine hydrolysis, degradation of polyurethane, and a heterogeneous catalytic reaction: methane coupling on titanium carbide. Our results demonstrated accurate predictions of reaction barriers for these reactions via our active learning scheme, with a notable reduction in the total number of configurations required to achieve comparable performance to other active learning methods.