I will discuss the capabilities of an open-source framework my group develops to apply deep networks to molecular structure and properties. NN-based models of potential energy surfaces (PESs) are one exciting application, and TensorMol enables the prediction of energies and forces inexpensively and accurately. We are also interested totally new chemical models which use the unique capabilities of machine learning answer new types of chemical questions. I will present models for directly predicting molecules with desired properties, and obtaining interesting trajectory and reactivity information without moving atoms at all.
Back to Workshop IV: Synergies between Machine Learning and Physical Models