ESPResSo Summer School: Prof. Dr. Christine Peter

Simulating biological soft matter across scales: making use of machine
learning methods

Coarse graining and combining particle-based models across scales has long been one important ingredient in overcoming the size and time scale limitations of purely atomistic approaches. In this context, linking the simulation scales and assessing and improving the inevitable shortcomings of the lower resolution models remains an ongoing effort in which machine learning (ML) plays an increasingly important role. Generally, in bottom-up coarse-graining, coarse-grained (CG) interactions are devised such that an accurate representation of a higher- resolution (e.g. atomistic) sampling of configurational phase space is achieved. Recently, traditional bottom-up methods have been complemented by machine learning (ML) approaches. ML methods can be used to derive or validate CG models by matching the sampling of a (relatively complex) free-energy surface as opposed to low- dimensional target functions/properties. For example, high-dimensional free energy surfaces can be extracted from atomistic simulations with the help of artificial neural networks (NN) – which can then be employed for simulations on a CG level of resolution. Secondly, ML methods can also be employed to obtain low- dimensional representations of the sampling of phase space or to identify suitable collective variables that describe the states and the dynamics of a system. This information can then be directly fed into the CG potentials or be employed to identify optimal CG representations and learn CG interactions. Moreover, the so-obtained low dimensional representations enable us to assess the consistency of the sampling in models at different levels of resolution, to go back and forth between the scales and ultimately to enhance and improve the sampling of the systems. In particular, they can be used as a basis for backmapping based enhanced sampling protocols.