(39f) Prediction of Atoms in Molecules with Matlab Graph Convolutional Network

Data-driven modeling methods is becoming more popular in chemical engineering and chemistry applications. Machine learning techniques are used for applications such as grouping similar equipment to detect outliers, fault detection and diagnosis, process inferential and optimization in chemical engineering space. Deep learning helps engineers with tasks such as identifying and tagging equipment, detecting defects, and developing predictive models to sort molecules based on their mass spectra data. Graph convolutional networks can be used directly on graphs and they utilize structural information. Molecules and social networks can be represented as graphs which makes them great inputs for graph convolutional networks. In a graph structure of a molecule, atoms in the molecules correspond to the nodes and chemical bonds correspond to the edges. With a graph convolutional neural network, types of unlabeled atoms in the molecule can be predicted from the molecules’ graph representations.

In this talk, we will share user stories on applications of data science to molecules and we will share a MATLAB workflow example of how graph convolutional networks can be used to predict the atoms in a molecule using QM7 data set.