Blake Duschatko1,Jonathan Vandermause1,Nicola Molinari1,Boris Kozinsky1
Harvard University1
Blake Duschatko1,Jonathan Vandermause1,Nicola Molinari1,Boris Kozinsky1
Harvard University1
Despite continuing advances in the development of interatomic potentials, biologically relevant long length- and time-scales remain inaccessible. Recently, machine learning has shown promise in developing sophisticated “bottom-up” coarse grained potentials, wherein the thermodynamic properties of the all-atom system are preserved. Despite the implications of successfully developing these models, a number of roadblocks remain. <br/> <br/>In this work, we demonstrate the use of Gaussian process (GP) regression in designing an on-the-fly active learning framework for coarse grained modeling. In particular, we show that the inherent uncertainty of GP’s allow these models to more seamlessly be transferred across chemically adjacent systems, addressing a persistent problem in the field. Moreover, we examine the applicability of this method to a variety of systems and explore ways in which uncertainty can serve as a probe to the fidelity of system property predictions in ways that traditional diagnostics, such as mean force errors, cannot.