Molecular mechanics force fields enable a wide variety of computational chemistry calculations, but also have well documented limitations.  Quantifying these limitations is nearly impossible due to the considerable human hours it takes to build a new force field. In an effort to help improve these essential models the Open Force Field Group (openforcefield.org) is developing a machinery that could automatically generate classical all-atom force fields. An important part of this process will be automating chemical perception -- the way force field parameters are assigned to a molecule based on chemical environment. We created the SMIRKS Native Open Force Field (SMIRNOFF) format which allows for changes to easily be made in both the chemical perception and quantitative parameter space. SMIRKS is a cheminformatics language for describing molecular fragments allowing the first iteration of this format, SMIRNOFF99Frosst, to cover as much chemical space as typical present-day force fields, with more than 10x fewer adjustable parameters (DOI:10.1101/286542). We have previously presented tools which automatically generated SMIRKS matching the chemical complexity in SMIRNOFF99Frosst (DOI:10.13140/RG.2.2.18704.10249). We can now also learn chemical perception from reference quantum mechanical data (not just an existing force field).  In implementing these tools, we were inspired by the process a scientist goes through when determining new chemical perception and took advantage of established machine learning techniques using the molecular graph. For a set of reference molecules and associated quantum mechanical data our tool learns how to cluster molecular substrctures for each valence term; where each cluster receives the same force field parameter. Next it, automatically generates SMIRKS patterns to match each cluster.  Here, we will present the automatically generated chemical perception necessary to parameterize the valence terms for a set of example molecules. This automated machinery will remove the necessity for a chemical expert to spend countless human hours assigning a force field’s chemical perception. The Open Force Field Group is also buiding tools to use Bayesian inference to learn the force field parameters which combined with our chemical perception machinery will automate the whole parameterization process.  This machinery will allow for improved force field science and in turn a better understanding of the results from simulations using them.

Automated chemical perception for force field parameterization

256th National Meeting (2018)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.