AI systems and human novices share a difficult problem: repairing incorrect models to improve expertise. For people, the use of analogies during instruction can augment the repair of science knowledge. Enabling AI systems to do the same involves several challenges: representing knowledge in commonsense science domains, constructing analogies to transfer knowledge, and flexibly revising domain knowledge. We address these issues by using qualitative models for representing knowledge, the Structure-Mapping Engine for analogical mapping, and a computational model of conceptual change for revising knowledge. In our simulation trials, we initialize the system with one of several student misconceptions of the day/night cycle from the cognitive science literature. The system automatically repairs these misconceptions using an analogy, expressed using natural language, by: (1) validating analogical correspondences via user feedback; (2) transferring knowledge from the base domain, and (3) constructing new explanations to repair misconceptions.