Steered molecular dynamics is a powerful technique for investigating reaction pathways that are otherwise difficult to access. However, the method’s reliability strongly depends on the choice of collective variables (CVs). Selecting appropriate CVs for a given system requires a deep understanding of its fundamental properties, because poorly chosen variables may fail to capture essential reaction pathways or produce unphysical outcomes. We propose a novel, intuitive method to train a soft actor–critic reinforcement learning agent to manipulate an ongoing trajectory by applying small, physically realistic forces (e.g., a thermal-bath–like perturbation) to reach goal states. Instead of predefined CVs, the agent uses one or multiple target molecular states as goals to guide the dynamics.
Alexander Scrimgeour