When designing a machine learning model for a predictive chemistry task, one of the first decisions that generally needs to be made is whether to employ a structure- or a descriptor-based representation. Descriptor-based representations usually require less complex model architectures and fewer data to yield good performance, but the selection of a suitable descriptor set is task-specific, and the evaluation of these descriptors for each data point can quickly become computationally expensive. Structure-based representations on the other hand are computationally inexpensive and not task-specific, but the associated models tend to require a lot of data to reach an acceptable accuracy.<br/>In this talk, we will discuss the recently proposed quantum mechanics-augmented graph neural network (ml-QM-GNN) architecture which aims to bridge these two representation strategies. In our hybrid approach, inexpensive structure-based representations are used in a first step to predict a set of atom- and bond-level descriptors derived from density functional theory calculations. These estimated QM descriptors are subsequently combined with the original structure-based representation to make reactivity predictions.<br/>Through consideration of a set of applications, we will demonstrate that our ml-QM-GNNs lead to significant improvements over analogous structure-based as well as descriptor-based models, not only in terms of overall accuracy but also in generalization to “unseen” compound classes. Even when training the model with only a couple hundred labeled data points, the ml-QM-GNN outperforms other state-of-the-art model architectures that have been applied to the same tasks.<br/>Furthermore, because the predictions of our model are grounded in QM descriptors, we can relate the predictions made to the traditional conceptual frameworks, commonly used to gain qualitative insights into reactivity phenomena. This effort results in a productive synergy between chemical theory and data science, wherein our QM-augmented models provide a data-driven confirmation of previous qualitative analyses, and these analyses in their turn facilitate insights into the decision-making process occurring inside ml-QM-GNNs.

Improving the performance, generalizability and explainability of neural networks for reactivity prediction through quantum chemistry-augmentation

Fall 2022 abstracts, posters, and presentations will launch in mid-September.

ACS Fall 2022

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.