<p>Digital twins of complex systems often involve the collaboration of multiple stakeholders. These domain experts do not speak the same domain languages, and this problem worsens when no domain-specific representations are usable by the twin. In controlled environment agriculture, digital twins must characterize crops via multiple concerns. That is, the digital twin must convey actionable results to its operators, who must assess the phenological and morphological development of the crop. Crop models are typically built to represent either phenology or morphology, making it difficult to simulate consistent representations across concerns. To address this, we introduce a model transformation based on a combinatorial optimization problem and study the applicability of genetic and reinforcement learning algorithms. We validate our methodology on the refinement of simulated strawberry phenological states and compare the applicability, performance, and diversity of our algorithms.</p>