To succeed in STEM, students need to learn to use visual representations. Most prior research has focused on conceptual knowledge about visual representations that is acquired via verbally mediated forms of learning. However, students also need perceptual fluency: the ability to rapidly and effortlessly translate among representations. Perceptual fluency is acquired via nonverbal, implicit learning processes. A challenge for instructional interventions that focus on implicit learning is to model students' knowledge acquisition. Because implicit learning is non-verbal, we cannot rely on traditional methods, such as expert interviews or student think-alouds. This paper uses similarity learning, a machine learning method that can assess how people perceive similarity between visual representations. We used this approach to model how undergraduate students perceive similarity between visual representations of chemical molecules. The approach achieved good accuracy in predicting students' similarity judgments and expands expert predictions of how students might perceive visual representations of molecules. [For the full proceedings, see ED592609.]