This study compared modeling skills and knowledge structures of four groups as seen in their understanding of air quality. The four groups were: experts (atmospheric scientists), intermediates (upper-level graduate students in a different field), advanced novices (talented 11th and 12th graders), and novices (10th graders). It was found that when the levels of modeling skills were measured, for most skills there was a gradual increase across the spectrum from the novices to the advanced novices to the intermediates to the experts. The study found the experts used model-based reasoning, the intermediates and advanced novices used relation-based reasoning, and the novices used phenomena-based reasoning to anticipate conclusions. The experts and intermediates used more bi-variable relationships in experimental design and anticipated conclusions, but used more multiple-variable relationships in identifying relationships. By contrast, the advanced novices and novices mostly used bi-variable relationships in all modeling skills. Based on these findings, we suggest design principles for model-based teaching and learning such as designing learning activities to encourage model-based reasoning, scaffolding one's modeling with multiple representations, testing models in authentic situations, and nurturing domain-specific knowledge during modeling.