Abstract
The optimization of chemical reactions requires navigating a high-dimensional design space composed of both discrete and continuous variables. Although one-hot encoding has been widely adopted, it lacks chemically meaningful information and suffers from sparsity and poor generalization. To address these limitations, we explored the use of pretrained molecular foundation models to generate latent representations as input variables for optimization. However, rigorously comparing different combinations of reaction representations and optimization algorithms remains a time- and resource-intensive challenge. In this work, we developed an end-to-end benchmarking platform that systematically evaluates diverse encoding schemes and optimization strategies under identical conditions. The platform automates the entire workflow from data preprocessing to result analysis, supporting fair comparison across multiple representation-optimizer combinations. Furthermore, we designed a custom reaction representation that integrates a 3D equivariant encoder with a bidirectional cross-attention module to explicitly capture interactions between reaction components. The proposed platform provides a scalable foundation for reaction optimization and advances the feasibility of autonomous experimental systems.