“Lubrication Brain” ― A machine learning framework of lubrication oil molecule design

In this work, we propose the concept of "Lubrication Brain". We adopt the Generative Adversarial Networks (GAN) coupling with reinforcement learning to automatically generate new molecules of Lubrication oil with desired properties. We pre-train a fully connected feedforward artificial neural network (NN) from experimental results to predict magnitude of properties of new molecules. This NN is embodied into GAN to evaluate the properties of new molecules, which serves as inputs of reinforcement learning to make GAN generate molecules with targeted properties. The application of "Lubrication Brain" on designing diester oil molecule with high flash point validates our approach which open new paradigm to design Lubrication oils.