The world faces several challenges and opportunities associated with new materials. For example, in the field of renewable energy, access to cheap battery materials could transform our energy systems and help save the planet from severe climate change. Over the past century, chemists have focused on the theoretical prediction of structure to property relationships for molecules and materials. For a given molecular library, screening for structure-property pairs can lead to the development of correlations that aid in molecular design. A much more challenging and in general, open problem, is the idea of inverse design, given the desired set of desired properties, generate a candidate material in silico for experimental testing. Generative models can help solve this challenge by providing compact and relatively low-dimensional latent spaces that can be employed to search for candidate structures never seen in the training of the models. In this talk, I will describe the general workflow for an automated, "self-driving" materials laboratory that can be helped by tools from machine learning. I will then focus on our recent work on autoencoders and development of generative adversarial networks for generating candidate molecules, and also on optimization algorithms for robotic machinery. I will end with a vision of the interesting open questions and opportunities.
I will also briefly mention our work on quantum computer machine learning. Namely, I will talk about our development the quantum mechanical autoencoder and a tool called the "quantum neuron" that allows for extending ideas and techniques. from deep learning to the quantum domain. I will also present a poster on the subject of quantum machine learning to talk about it with the interested workshop participants.
Back to New Deep Learning Techniques