Machine Learning Innovation in Support of Chemistry and Materials Science

Описание: Session II - Machine Learning Innovation in Support of Chemistry and Materials Science of FRP Symposium Advancing Chemical and Materials Science through Machine Learning

Frank Noe, Professor, Department of Mathematics and Computer Science, Freie Universität Berlin

Deep learning for molecular sciences

Abstract: AI, and specifically deep ML methods have a profound impact on industry and information technology. But since recently AI methods are also changing the way we do science. In this talk I will present some of our recent efforts to build machine learning methods that attack fundamental problems in physical and chemical sciences: the sampling problem in physical many-body systems, and the solution of the quantum-chemical electronic Schrödinger equation. Key in making progress in these hard problems with ML is to interrogate the physical system about what the learning problem should be, and to encode physical structures, such as symmetries and conservation laws, into the ML model.

Risi Kondor, Associate Professor, Departments of Computer Science and Statistics, The University of Chicago

Equivariant neural neural networks for physics and chemistry

Abstract: Unlike many other “big data” domains, when using machine learning in physics or chemistry, we have physical laws that the learning algorithm must satisfy, such as invariance to translations, rotations, and other symmetries. The theory of group equivariant neural networks, built on representation theory, provides a general framework for incorporating such symmetries in a computationally efficient manner. I will discuss this general theory, highlight some recent practical successes and describe some future directions.

Bidisha Samanta, Research Engineer, Google

NeVAE: A deep generative model for molecular graphs

Abstract: Deep generative models have been praised for their ability to learn smooth latent representation of images, text, and audio, which can then be used to generate new, plausible data. However, current generative models are unable to work with molecular graphs due to their unique characteristics—their underlying structure is not Euclidean or grid-like, they remain isomorphic under permutation of the nodes labels, and they come with a different number of nodes and edges. In our work we propose a variational autoencoder for molecular graphs, whose encoder and decoder are specially designed to account for the above properties by means of several technical innovations. We further develop a gradient-based algorithm to optimize the decoder of our model so that it learns to generate molecules that maximize the value of a certain property of interest and, given a molecule of interest, it is able to optimize the spatial configuration of its atoms for greater stability.

Panel Discussion Moderated by Brian Kulis, Associate Professor. Electrical and Computer Engineering, Boston University