Shared representations for model-driven reaction optimisation

Determine what a suitable shared representation for chemically diverse reactions should contain in order to support model-driven, multi-objective reaction optimisation across heterogeneous reaction components.

Background

Model-driven reaction optimisation depends on representing chemically diverse components—including ligands, catalysts, bases, solvents, and precursors—in a form that enables meaningful comparison and prediction. Molecular descriptor libraries are often restricted to particular component classes and may not transfer across chemically distinct systems, whereas one-hot encodings do not capture chemical similarity. The paper introduces dynamically learned language-model embeddings as a practical approach to this representation problem, but the general question of what information a shared representation should contain across heterogeneous reaction spaces remains unresolved.