Demonstrates a method to couple robust physics models with data driven machine learning models - getting the best of both worlds.
Investigated how two different approaches to featurizing and fingerprinting a host-guest system and the ultimate impact upon accuracy. Looked at fingerprinting the structure using a modern Extended connectivity fingerprint and into a deep neural network. In another model, investigated allowing a graph convolutional network produce a fingerprint and fed that fingerprint into densely connected linear layers. In both approaches, coupling physics from a endpoint model resulted in an accuracy improvement!