Artificial intelligence meets molecular dynamics: Certified machine learning for resolution transformation

Coarse-graining of molecular dynamics simulations is a common practice, allowing scientists to describe larger simulation systems over longer time periods. During coarse-graining of the system, groups of atoms are replaced with particles carrying the chemical-physical properties of the underlying atoms. This process is necessarily accompanied by a loss of structural information on the stereochemistry and the relative orientation of the individual chemical groups, making the conversion to a higher-level resolution a non-trivial operation. In this project, we will develop a novel resolution transformation methodology enabling conversion from coarse-grained back to all-atom resolution using a physics-informed machine learning approach. Our approach avoids systematic errors on a methodological level, which are inherent to existing methods in the literature, and speeds up the computation. The explicit encoding of the chemical and physical properties within our machine learning approach robustifies and speeds up the training with fewer data and better approximation quality. Besides, this approach allows the development of residual-based a posteriori error indicators to certify the surrogate models.

This project is funded within the Artificial Intelligence Software Academy, funded by the Ministry of Science, Research and the Arts of Baden-Württemberg