Fig. 1: Reaction yield prediction framework for small-scale data.

A Traditional yield optimization. B The overall framework of our model with RS-coreset. Our yield prediction result can be achieved via an iterative procedure, where each iteration includes 3 steps. Step 1 (“yield evaluation”), evaluate the yields of selected reactions by chemists. Step 2 (“representation learning”), update the representation learning model with the newly added experimental data. Step 3 (“data selection”), select the most informative reactions in the reaction space guided by our coreset method. After several iterations, the representation for reaction space becomes stable and then we can build the final yield prediction model upon it.