(106d) Improving the Reusability of Machine Learning Models for Materials

Machine learning models have proven useful in a variety of chemical and materials applications. However, one drawback of these models is that they are often geared to a particular application, and thus a new model must be created for a new application. This is particularly noticeable in applications that feature guest-host interactions, broadly defined, as there is a huge combinatorial challenge in considering many possible guests and many possible hosts. For example, this combinatorial challenge applies to catalysis, adsorption, doping, and intercalation. In this work, we develop and apply strategies to improve the generality and reusability of machine learning models.

We have developed a machine-learning architecture that takes advantage of inherent aspects of chemical systems, and is particularly geared towards those that feature guest-host interactions. This framework partially decouples the guests from the hosts by introducing intermediate variables, which are a set of implicit host properties that control how the host interacts with all guests. These intermediate variables are learned during the fitting process. We create separate sub-models for different host elements for predicting these intermediate variables, and separate sub-models for how different guests respond to these host variables. The sub-models are all fit simultaneously. Our framework takes advantage of the fact that elements are discrete entities, and greatly simplifies the huge combinatorial challenge of considering many possible guests and many possible hosts. We apply this framework to multiple datasets and show how it can be used to improve the efficiency and accessibility of initial screening in a variety of applications.