Navigating Transition Path Sampling with Machine Learning Potentials: Insights and Challenges

In the realm of molecular dynamics, capturing elusive transitions and unveiling potential energy landscapes is a paramount pursuit. Reactive events that lead a system from state A to state B through a transition path are often masked by an ensemble of energetically similar pathways. Transition path sampling (TPS) emerges as a potent tool to uncover these paths, although their infrequency in simulations due to high energy requirements and limited statistical occurrence remains a hurdle.<br/>Machine learning potentials offer a promising avenue to model and discern these intricate transition paths, opening the door to new possibilities. The question that arises is: how do we select the right model and data?<br/>Diverse machine learning potentials, such as HIPNN and ANI, have been shown to yield varying transition paths during sampling. An illustration is found in our exploration of alanine dipeptide, a system replete with multiple dihedral degrees of freedom. Intriguingly, even when accuracy metrics like mean absolute error (MAE) and root mean square error (RMSE) align, these static metrics assessed on a portion of the dataset may not suffice to identify the superior model. They serve as preliminary benchmarks, rather than comprehensive tests of model performance.<br/>One avenue toward superior performance emerges through active learning sampling in the transition region. Our demonstrated approach significantly bolsters accuracy, reducing energy errors to a remarkable 0.5 kcal/mol. However, the path to progress is paved with caution when selecting transition paths for machine learning model testing. We emphasize this point using the example of azobenzene, a seemingly uncomplicated system with limited degrees of freedom, yet it poses a significant challenge for electronic structure calculations. Machine learning results may appear in perfect alignment with easily calculable paths, but this can be misleading if the unique open-shell nature of the transition state is overlooked. Establishing reliable benchmarks for machine learning, particularly in the context of transition path sampling, remains an intricate and ongoing endeavor.<br/>Armed with these lessons of caution, our aim is to cultivate a deeper understanding of how to assess the accuracy of machine learning models in dynamical processes.