Transition-state search algorithm based on deep reinforcement learning

Abstract We present a transition-state (TS) search algorithm based on deep reinforcement learning (DRL) that inherits two important features, namely the use of the DQN algorithm and the use of curvatures of the potential energy surfaces (PESs), from contemporary RL theory and computational chemistry, respectively. Because this is, to our knowledge, the first attempt to apply a hybrid DRL method using curvatures to the straightforward TS search task, we focus on two points. The first one is whether our DRL method works for the TS search task. The second point is the transferability of the trained DQN model to other PESs that are not used in the training processes. The computational results are discussed in relation to future applications to molecular systems.