Citations in front:
Software Demo:
Features or descriptors are the crucial components in generating reliable and accurate machine-learning models.
Featurization is the key step in any machine learning (ML) routine. In mathematical terms, the y-block is a property or class that we focus on and try to predict, and the x-block is a list of features that, as we believe, might govern the property or class. It is important to understand that the x-block itself might not contain the answer to the problem, but a well prepared x-block that helps to segregate classes (classification problem) or find the trend (regression problem) might reveal a meaningful relationship between the x-block components, which could be a solution to a given problem. (Feature Engineering)
Regardless of the ultimate goal behind why we are using ML, the result strongly depends on how the x-block is generated and what features we use. Nowadays, we know that property depends on composition and structure (All the descriptions are around these two to expand), and we try to use both to rationally design new material or rationalize our discovery. Working with crystallographic databases, we have access to structure-derived features (atomic distances, symmetry, coordination geometry, etc.) and composition-derived features. In materials ML, the more structure features included, the more advanced and specific the ML model will be. But there is a nuance. The advantage of including structure-derived features is completeness of the description; however, the disadvantage is the mandatory structural information that has to be specified before the prediction could be made for an arbitrary composition (Figure 1).
The problem is that we cannot know ahead of time the structure of a new material that we try to predict, therefore cannot include the structure-derived features. To overcome this obstacle and build effective ML models there are two approaches: 1) use the most complete and meaningful list of elemental properties and create a model based on compositional information only; 2) try to guess structure-based features by using a separate crystal structure classification model to predict the most probable structure based on composition, and utilize the predicted structure as the basis for the structure-based features. Most likely, a set of features based only on composition, might not perform as good as the set with composition- and structure-based features together. However, striving for the best composition-based set is always worth it, given that the composition still is one of the most specific descriptions of a given compound.
In the current study, we demonstrate the application of the proposed property list to tackle the crystal classification problem and validate it with the prediction and synthesis of a new compound. We propose one of the best-performing lists of properties of elements that could be used to generate the set of features for successful ML predictions.The Oliynyk list is based on the properties used in numerous lab-tested ML studies (described later in the text) and was confirmed to perform the best compared to common featurizers for the materials data sample sizes of tens-hundreds-thousands of data points (Figure 2).
Figure 2. Performance of Oliynyk property list compared to common featurizers.
The proposed list of elemental properties can serve as the basis for featurization, containing the most of properties, some of which were gathered from various sources, and the others were predicted by our ML model (described by them), in case some values were missing for a particular element.