Evidence That Less Can Be More for Transferable Force Fields.

Graph-based parameter assignment has been the basis for developing transferable force fields for molecular dynamics simulations for decades. Nevertheless, transferable force fields vary in how specifically terms are defined with respect to the molecular graph and the procedures for generating parametrization data. More-specific force-field terms increase the complexity of the force field, theoretically increasing accuracy but also increasing training data requirements. In contrast, less-specific force fields can be reused across larger regions of chemical space, theoretically reducing accuracy but also reducing the number of parameters and training data requirements. Here, the tradeoffs between force-field specificity and accuracy are quantified by parametrizing three new sets of force fields with varying levels of graph specificity, using a shared procedure for generating training data. These force fields are benchmarked for their ability to reproduce the structural features and liquid properties of 87 organic molecules at 146 distinct state points. The overall accuracy for properties that were directly trained on rapidly saturates as the graph specificity of the force-field increases. From this, we conclude there is at best a marginal benefit of using less transferable and more complex force fields with common sources of quantum-chemically derived training data. When looking at properties unseen during training, there is some evidence that the more-complex force fields even perform slightly worse. These results are rationalized by the fortuitous regularization of force fields based on less-specific and more-transferable atom types. Both the saturation in the accuracy of training properties and the marginally worse performance on off-target properties fundamentally contradict the expectation that bespoke force fields are generally more accurate, given their larger number of parameters, and suggests that increasing force-field complexity should be carefully justified against performance gains and balanced against available training data.