State predictive information bottleneck

The ability to make sense of the massive amounts high-dimensional data generated from molecular dynamics (MD) simulations is heavily dependent on knowledge a low dimensional manifold (parameterized by reaction coordinate or RC) that typically distinguishes between relevant metastable states and which captures slow interest. Methods based machine learning artificial intelligence have been proposed over years deal with such low-dimensional manifolds, but they are often criticized for disconnect more traditional physically interpretable approaches. To concerns, in this work, we propose deep State Predictive Information Bottleneck (SPIB) approach learn RC high simulation trajectories. We demonstrate analytically numerically how learnt deeply connected committor chemical physics, can be used accurately identify transition states. A crucial hyperparameter time-delay, far into future algorithm should predictions about. Through careful comparisons benchmark systems, choice gives useful control coarse-grained want state classification system be. thus believe work represents step forward systematic application ideas way bridges gap physics.