Most current AI systems for molecular generation function like black-box translators; they map inputs to outputs without an explicit intermediate structure. We propose a different paradigm:
viewing molecular design as a compiler pipeline for chemistry.
Graph rewrites serve as the intermediate representation, and statistical mechanics, through energy functions and the partition function, provides the optimization layer
that guides which transformations are most favorable.
This approach enables the construction of hybrid systems that intelligently balance
physical reasoning with machine learning, offering greater interpretability and control
than purely data-driven generative models.
