Models are common in chemistry. When these models can be described mathematically, their real world implications can often be simulated using computers, enabling the use of more complex models in hopes of improving scientific predictions. Prior to providing useful results these models must often be calibrated against existing scientific data. Separately, machine learning has recently gained significant traction in many applications. The algorithms underpinning machine learning often similarly require calibration prior to application. This work provides mathematical and numerical results connecting these two areas. Specifically, we consider novel applications of classification when creating molecular models such as those used in coarse-grained molecular dynamics. We focus on the concept of adversaries, a tactic that has recently gained traction in the machine learning community, and use this framework to analyze the difference between various coarse-grained ensembles and to parameterize new coarse-grained force-fields. Collectively, we show that classification is an effective tool for understanding the differences between high dimensional free energy surfaces and that adversarial parameterization strategies are theoretically and numerically feasible for coarse-grained models.