Supervisor: Jacob Sherson

Proposing new materials or molecules with desirable properties is a never ending quest in material science and chemistry. Previously this task was primarily handled by the researcher, but with the advent of machine learning this task has partly shifted to the machine.

Generative Adversarial Networks (GANs) is one such model that creates new samples based on previous observations.  In this framework one network learns to produce realistic data by competing with another network that tries to distinguish generated data from real data.

In this colloquium GANs will be demonstrated initially in the context of image generation, in which they were first employed. Then a GAN will be utilized to propose new molecules by training on a database of small organic molecules. It will be demonstrated how one can learn to not only mimic the samples in the database, but also propose novel molecules by introducing reinforcement learning to bias the samples towards features such as druglikeness.