From prompt engineering to modeling: secondary science teachers' use of generative AI to engineer sweetener molecules

Scientific modeling is a core epistemic practice in chemistry, yet supporting learners in visualizing and making sense of abstract molecular structures and functions remains a challenge. While generative artificial intelligence (GenAI) offers new representational capabilities, it remains unclear how text-based prompt engineering intersects with established theories of scientific modeling. This study addresses this gap by reconceptualizing prompt engineering as a form of molecular modeling within AIMS, a chemistry-specific GenAI-supported molecular modeling environment. Through a qualitative case study, 11 secondary science teachers used GenAI embedded in AIMS to design novel sweetener molecules. Applying the CLEAR and REFINE analytic frameworks through the lens of Halloun's Modeling Theory, the study examined how participants translated disciplinary knowledge into iterative prompt-based molecular designs. The analysis concludes with three key findings. Participants' prompts functioned as externalized models that embedded assumptions about molecular structure-function relationships and design constraints within explicit chemical and theoretical requirements. As iterations progressed through model testing, reasoning shifted from broad appeals for scientific accuracy to precise, mechanistic constraints, including spatial arrangements and molecular properties. Revisions operated as controlled experiments where participants tightened or relaxed constraints to test the plausibility of GenAI outputs. These findings advance chemistry education by demonstrating that a chemistry-specific GenAI-supported modeling tool can function as a modeling assistant rather than an answer engine, scaffolding the core epistemic practices of construction, evaluation, and refinement in chemistry education. This reconceptualization positions prompt engineering as a source of written evidence for how learners articulated, tested, and revised chemical constraints, offering educators new ways to support and assess student modeling practices.