Interpretable Machine Learning in Solid-State Chemistry, with Applications to Perovskites, Spinels, and Rare-Earth Intermetallics: Finding Descriptors Using Decision Trees

Machine-learning methods have exciting
potential to aid
materials
discovery, but their wider adoption can be hindered by the opaqueness
of many models. Even if these models are accurate, the inability to
understand the basis for the predictions breeds skepticism. Thus,
it is imperative to develop machine-learning models that are explainable
and interpretable so that researchers can judge for themselves if
the predictions are consistent with their own scientific understanding
and chemical insight. In this spirit, the sure independence screening
and sparsifying operator (SISSO) method was recently proposed as an
effective way to identify the simplest combination of chemical descriptors
needed to solve classification and regression problems in materials
science. This approach uses domain overlap (DO) as the criterion to
find the most informative descriptors in classification problems,
but sometimes a low score can be assigned to useful descriptors when
there are outliers or when samples belonging to a class are clustered
in different regions of the feature space. Here, we present a hypothesis
that the performance can be improved by implementing decision trees
(DT) instead of DO as the scoring function to find the best descriptors.
This modified approach was tested on three important structural classification
problems in solid-state chemistry: perovskites, spinels, and rare-earth
intermetallics. In all cases, the DT scoring gave better features
and significantly improved accuracies of ≥0.91 for the training
sets and ≥0.86 for the test sets.