Machine Learning Physics-based Chemical Models

Purpose: atomistic models of chemical and biological systems are either computationally expensive or not accurate enough to make reliable predictions. In addition, systematic errors in predictions due to uncertainties in the model are typically neglected. Aim: I will solve these problems by building new, force fields (FFs) based on well-established physics.

Parameters with well-defined uncertainties will be trained on big data sets from quantum chemistry, using our recently released Alexandria Chemistry Toolkit. How: Development of FFs will be done by a postdoc and a graduate student under my supervision. Preliminary results strongly suggest that working versions will be available within two years.

Further refinements and new applications will follow.

Importance: The new FFs will be chemistry-aware in that they can distinguish isomers by energy, which will allow for countless new applications such as studying pH dependent processes and drug design. We will explore many of those new possibilities with other groups in local and international research collaborations.

Implementation in OpenMM and GROMACS and dissemination through our widely used website at virtualchemistry.org will speed up adoption of the FFs.

Breakthrough potential: Systematic comparison of models will speed-up drug and materials design, correct uncertainty estimates will reduce computational cost and, e.g. accurate predictions of electrolyte properties will aid battery development.