Data Mining Approaches to High-Throughput Crystal Structure and Compound Prediction

Abstract

Predicting unknown inorganic compounds and their crystal structure is a critical step of high-throughput computational materials design and discovery. One way to achieve efficient compound prediction is to use data mining or machine learning methods. In this chapter we present a few algorithms for data mining compound prediction and their applications to different materials discovery problems. In particular, the patterns or correlations governing phase stability for experimental or computational inorganic compound databases are statistically learned and used to build probabilistic or regression models to identify novel compounds and their crystal structures. The stability of those compound candidates is then assessed using ab initio techniques. Finally, we report a few cases where data mining driven computational predictions were experimentally confirmed through inorganic synthesis.

Electronic properties of reduced molybdenum oxides

“…ΔH per cation is shown with oxygen stoichiometry in Figure 6, in which the solid line indicates the convex hull formed by the most stable DFT ground state structures for a given stoichiometry. 70 The convex hull connects the phases that are stable against decomposition into the other phases, and consists of MoO 51,71 This discrepancy can be related to the difference in comparing 0 K ground state structures to finite temperature experimental data, and effects of entropy as well as errors innate to the DFT method can be considered. Firstly, there is a lack of data on the low temperature stabilities of molybdenum oxides.…”

A general-purpose machine learning framework for predicting properties of inorganic materials

“…To date, materials scientists have used machine learning to build predictive models for a handful of applications. [14][15][16][17][18][19][20][21][22][23][24][25][26][27][28] For example, there are now models to predict the melting temperatures of binary inorganic compounds, 22 the formation enthalpy crystalline compounds, 15,16,29 which crystal structure is likely to form at a certain composition, 6,17,[30][31][32] band gap energies of certain classes of crystals, 33,34 and the mechanical properties of metal alloys. 25,26 While these models demonstrate the promise of machine learning, they only cover a small fraction of the properties used in materials design and the datasets available for creating such models.…”

Applications of Artificial Intelligence and Machine Learning Algorithms to Crystallization

“…Particularly, with the advent of real-time in situ monitoring of crystallization via process analytical technologies (PATs) − and the resulting generation of large high dimensional data sets (e.g., time series of spectra or particle images), statistical models based on machine learning can accurately describe in real-time, otherwise unattainable, solution and solid-state properties during crystallization . Input-output ML models have also been used to simulate and control the complex nonlinear crystallization dynamics and to relate critical quality attributes of the solid products with adjustable process input parameters. , In the polymorphism and crystal structure prediction (CSP) area, , which constitutes one of the enduring challenges in physical and chemical sciences, machine learning , and data-mining techniques have been proven to accelerate discovery of new crystalline materials , (including salts, solvates, and cocrystals) and structures, thereby saving tremendous experimental effort associated with labor-intensive experimental solid form screenings. Moreover, machine learning has contributed significantly to the in-silico prediction of properties that govern the behavior of crystalline materials, such as solubility and melting point, using as inputs a variety of molecular descriptors. , Finally, in high-throughput experimentation, usually undertaken during process development for complex organic biomacromolecules such as proteins and oligonucleotides, ML classification and clustering together with image analysis methods can quickly identify promising crystallization conditions, automating the development of emerging therapeutic modalities.…”

“…18 Input-output ML models have also been used to simulate and control the complex nonlinear crystallization dynamics and to relate critical quality attributes of the solid products with adjustable process input parameters. 19,20 In the polymorphism and crystal structure prediction (CSP) area, 21,22 which constitutes one of the enduring challenges in physical and chemical sciences, machine learning 23,24 and data-mining 25 techniques have been proven to accelerate discovery of new crystalline materials 26,27 (including salts, solvates, and cocrystals) and structures, thereby saving tremendous experimental effort associated with labor-intensive experimental solid form screenings. Moreover, machine learning has contributed significantly to the in-silico prediction of properties that govern the behavior of crystalline materials, such as solubility and melting point, using as inputs a variety of molecular descriptors.…”