Recently, Google AI and Caltech researchers discovered new metal oxides using ink printers and machine learning methods. Several studies employed machine learning to forecast material qualities, but the approaches were intrinsically specialised and failed to grasp the problem’s global nature. Let’s have a look at how it all went down.

Use of ML Algorithms

Machine learning algorithms were used to determine a material’s suitability for a given attribute. However, they have been used to narrow down the list of outstanding materials for every given attribute. As a result, researchers are averse to using machine learning to identify materials with particular features. Additionally, chemists frequently allege an ‘intuition’ regarding the patterns of reagent properties and composition ratios that govern material creation. If such patterns exist, they can be uncovered using data mining techniques when given a database of successful and unsuccessful reactions.

Here, the researchers have created two data science approaches for material discovery in high-order composition regions. In addition, the present research details the formulation and construction of the high-throughput process that feeds these systems data, as well as an illustration of a use case for leading discovery. The major discovery is that well-developed data science models can infer the phase behaviour of complicated materials using datasets that are not typically used for phase characterisation. Furthermore, these conclusions provide scientific value to the current datasets and aid in the development of new materials.

Advantages 

The application of data science-driven evaluation of experimental data complements quantum mechanical and machine learning-based phase prediction. Additionally, the detection of novel systems and compositions by optical data modelling can serve as a springboard for further research into new phases and/or assessing that three-cation compositions show unique features. Moreover, the approach is founded on the principles of combinatorial materials science, where the synthesis of mixture libraries is linked to the measurement of desired attributes. While this approach provides a direct path to discovering a desirable property in a particular composition system, it constrains the discovery of many composition areas due to the increased relative cost of property measurements and the requirement to measure each composition library for each desired property.

Other Research Contributions

• The researchers in the United States employed typical RFs to forecast structure energies based on Voronoi tessellations and atomic characteristics. (Read here)
• The Chinese researchers predicted stable full-Heusler compounds using convolutional neural networks and transfer learning. (Read here)
• A team of researchers from the United States employed a variety of regression and classification techniques to analyse a dataset of cathodes for increased solid oxide fuel cells. (Read here)
• Canadian researchers used cluster resolution feature selection to classify binary crystal formations. (Read here)
• Similarly, researchers in the United States use VAEs to predict crystal structures. (Read here)

Conclusion

Traditionally conducted ML research explores only one property at a time. As a result, the evaluation must be conducted multiple times on the same material. In addition, training data sets are particularly scarce in high-order composition fields, which allow for the tuning of many attributes via the development of a phase containing all three cations. However, the scarcity of training data for materials frequently limits prediction accuracy, particularly in composition areas for which no training data exists. Moreover, a handful of great reviews have been published on machine learning in materials science in general, as well as on machine learning in the chemical sciences, which might signal towards opening the road for new ventures.