Beneath Surface Similarity: Large Language Models Make Reasonable Scientific Analogies after Structure Abduction