Machine learning descriptors in materials chemistry: prediction and experimental validation synthesis of novel intermetallic UCd3

Abstract

Materials informatics uses data-driven approaches for the study and discovery of materials. Features or descriptors are the crucial components in generating reliable and accurate machine-learning models. While general data can be acquired through public and commercial sources, features must be tailored for a specific application. Common featurizers are suitable for generic chemical problems, but may not be ideal for solid state materials. Here, we have assembled the Oliynyk property list for feature generation which works well on limited datasets (50 to 1,000 training data points) in solid state materials domain. We applied Gaussian process regression to extrapolate and predict missing values in the element property list. Complete data in the feature list allows researchers to use any methods that require a complete x-block without any data gaps, such as SVM. To validate our updated property list and generated features based on this list, a classical crystallographic problem of classifying structure type was solved. Similarly to the radius ratio rule (Linus Pauling, 1929) and structure maps (Villars, 1983; Pettifor 1984), we demonstrate how 1:3 stoichiometry structure types could be classified with SVM method and the x-block based on the proposed list of properties. We validate the ML model experimentally by synthesizing a novel intermetallic UCd3.

Machine learning descriptors in materials chemistry used in multiple experimentally validated studies: Oliynyk elemental property dataset

“…The features had 63 properties of elements, including different size scales (obviously without Miracle), electronegativities, and physical properties, and DFT-generated data. The algorithms included [ 1 ] Support-vector machines (SVR), [2] Decision Tree, [3] Multi-layer Perceptron neural network (MLP-NN), [4] Cubic spline interpolation, and [ 5 ] Gaussian process regression (GPR). The root mean squared error (RMSE) values between the actual and predicted Miracle values were compared.…”

“…Here, we introduce the Oliynyk property list, a carefully curated Excel file containing 98 elemental features for atomic numbers from 1 to 93. This dataset has been used as the basis for featurization in various lab-tested studies [1] . If the properties had missing values, we predicted them to complete the dataset.…”

“…The information includes atomic number, period, group number, family number, metal/nonmetal classification, Mendeleev number, and l quantum number. Each family number corresponds to the following group: [1] alkali metals, [2] alkaline-earth metals, [3] f-block metals, [4] d-block metals, [5] p-block metals, [6] p-block metalloids, [7] p-block nonmetals, [8] chalcogens, and [9] noble gases.…”

Experimentally validated machine learning predictions of ultralow thermal conductivity for SnSe materials

“…Besides ESF and EDF, the construction of weighted element-to-chemical property features was carried out using a separate dataset -the Oliynyk elemental descriptors (OELDs). 46 OELDs contain detailed information on the chemical and physical properties of elements in the periodic table. These weighted element-to-chemical property features are composed of two groups:…”

Not as simple as we thought: a rigorous examination of data aggregation in materials informatics

“…Also falling under this category are methods that leverage domain knowledge to design better data features, more commonly known as 'feature engineering'. 17,18 More data: in this branch of the data-centric approach, the attention is shied to increasing the number of data points. This is generally considered to be more signicant in view of a better-performing statistical model 19,20 and a compelling alternative to vast domain knowledge.…”