Abstract:

The application of machine learning to chemistry has seen tremendous success in recent times. However, in most cases the applicability of the trained models from such studies is limited to points similar to training data, with far-off and borderline data points being extremely prone to errors. To handle such error-prone data points alternative approaches must be adopted. In an attempt to automate this scenario Guan and co-workers propose an automated workflow [1] which triggers expensive alternative computations, when the model is uncertain about its predictions. The work not only achieves impressive accuracy on benchmarks but also lays down the foundation for future work in uncertainty aware prediction. In this seminar, I will discuss the foundational aspects of uncertainty quantification implemented by the workflow, and how they can be used to obtain uncertainty aware predictions from a machine learning model.

1. Guan Y.; Lee T.; Wang K.; Yu S.; McWilliams J.; J. Chem. Inf. Model. 2023, 63, 12, 3751-3760