Evidence of Information Limitations in Bottom-Up Coarse-Graining Models

Chemically specific coarse-grained (CG) models exhibit simplified configurational phase spaces and hence, can potentially capture processes that occur on time and length scales that are too costly for direct atomistic simulation. However, there are fundamental and practical problems associated with coarse-graining, such as the incomplete understanding of CG errors in comparison with atomistic simulations, the lack of transferability of typical CG models to new chemistries, and the costly bespoke approach to developing new CG models from scratch. In this talk, I will present recent results on an automated methodology to parameterize CG models from quantum chemistry calculations. Using the throughput of this methodology, we have systematically characterized the sources of error in common bottom-up parameterization methods as a function of dimension reduction for over 50 independently trained CG models. We find clear evidence that these models are systematically information limited, rather than representability limited, which suggests further improvement is obtainable without resorting to more complex functional forms. Additional implications of these findings, as well as the conditions under which representability related errors arise will also be discussed.