Machine learning techniques are increasingly used by the computational chemistry community to predict molecular properties. Compared to traditional methods, which can be computationally prohibitive, machine learning models can predict certain properties with high accuracy and in much less time. One such machine learning model is a Graph Convolutional Network (GCN), a neural network architecture for performing machine learning on graphs. A Physics-Informed Neural Network (PINN) is a neural network in which physical laws are employed to guide the learning process. This talk will discuss how to combine the representational power of a GCN with the physical knowledge incorporated into a PINN to predict target molecular properties such as partial charges and atomization energies.
