Skip to main content
Event Date and Time
Roberto Covino, Frankfurt Institute for Advanced Studies, Frankfurt, Germany
I will present an artificial intelligence (AI) agent that learns molecular mechanisms from computer simulations. The agent integrates path theory, transition path sampling (TPS), deep learning, and symbolic AI to sample complex molecular self-organization phenomena and learn how to predict their outcome. TPS is a Markov Chain Monte Carlo method to sample the rare trajectories connecting metastable states. Using reinforcement learning, the agent iteratively trains a deep neural network on the outcomes of TPS simulation attempts. In this way, it increases the rare-event sampling efficiency while gradually revealing the underlying mechanism. The AI agent learns the committor function of the rare event encoded in the trained neural network. Symbolic regression distills quantitative models that reveal human-understandable mechanistic insight. With this algorithm, we studied the dissociation of ion pairs in solution, the nucleation of methane clathrates, and the spontaneous assembly of a model transmembrane. In all three cases, the AI agent produces quantitative mechanistic models that capture the complex molecular reorganizations occurring during the transitions. For the transmembrane assembly, in less than 20 days and with minimal human intervention, the AI agent accumulates a total of 5 ms Martini simulation time distributed over 10000 trajectories, collecting approximately 4000 unbiased transition paths. In conclusion, our AI enables the sampling of rare events by autonomously driving many parallel simulations with minimal human intervention and aids their mechanistic interpretation.