The goal of this project is to assist chemists in finding new materials for battery cathodes, by training a generative model that produces promising candidate molecules. Our training data is an existing chemical database of several thousand organic molecules. Each molecule has been annotated with chemical information such as: how much charge can it hold? can it be easily synthesised? is it sufficiently stable? We use this data to train two neural networks: 1. A predictive model that, given a description of the molecule, estimates its chemical properties. 2. A generative model, trained by reinforcement learning on the predictive one, that creates promising cathode molecules. A further goal is to integrate domain-specific knowledge from chemists into the model. For example, a chemist may have a certain idea of the structure of molecule they want, and we then want to bias the model towards generating this type of molecules. Also, chemists have defined many kinds of "molecular descriptors" - symbolic molecular properties that can be easily computed. The model should learn which descriptor values are relevant, and how to generate molecules with "good" descriptor values. A major question in our project is how to integrate domain-specific and symbolic information into a generative model.