Interpretable Graph-Network-Based Machine Learning Models via Molecular Fragmentation

Chemists have long benefitted from the ability to understand and interpret the predictions of computational models. With the current shift to more complex deep learning models, in many situations that utility is lost. In this work, we expand on our previously work on computational thermochemistry and propose an interpretable graph network, FragGraph(nodes), that provides decomposed predictions into fragment-wise contributions. We demonstrate the usefulness of our model in predicting a correction to density functional theory (DFT)-calculated atomization energies using Δ-learning. Our model predicts G4(MP2)-quality thermochemistry with an accuracy of <1 kJ mol^-1 for the GDB9 dataset. Besides the high accuracy of our predictions, we observe trends in the fragment corrections which quantitatively describe the deficiencies of B3LYP. Node-wise predictions significantly outperform our previous model predictions from a global state vector. This effect is most pronounced as we explore the generality by predicting on more diverse test sets indicating node-wise predictions are less sensitive to extending machine learning models to larger molecules.