In today’s world, numerous metal oxides are present which can be put to use in a wide variety of applications as catalysts, adsorbents, and semiconductors, to name a few. Multiple properties of these materials are relevant to existing technologies and can assist in developing new ones. Despite their critical functionalities, their potential is yet to be tapped as their number, and the extent of valuable properties is unexplored.
Conventional research only considers one property at a time during the experiments. Therefore, the evaluation is required to be repeated several times on the same material. Machine Learning methods were used in several studies to predict material properties, but the models were inherently specialized and failed to capture the universal nature of the problem. At present, there is a need to explore a large swath of chemical space as efficiently and quickly as possible.
Hence, in collaboration with the California Institute of Technology, Google AI has developed new synthesis and characterization methods to accelerate the discovery of such complex metal oxides with the help of Machine Learning (ML) techniques to screen out potential materials rapidly. Previously ML models were used to find out a suitable material with a given property. However, this time they have been used to shortlist exceptional materials for any particular property.
A significant challenge in realizing this strategy is the enormous search space one has to deal with. Generating novel compounds simply by considering various combinations would not result in a new crystalline structure. Instead, this approach would result in a combination of existing structures, and the whole idea of finding core single structure compositions stands defeated. Hence, the ML-based inferences would be used to find new structures.
To synthesize metal films, a customized inkjet printer has been used to print samples with different ratios of metals. The printing technology has been designed in such a way that it can dissolve metal nitrates or metal chlorides into the ink solution to print out various metals. A series of line segments are printed out on a glass plate, and numerous unique combinations can be engendered on each piece of glass. These boards are further baked to oxidize the metal, and the desired number of copies for each composition is printed. This method has emerged a hundred times faster than the traditional vapor deposition technology.
As the speed of sample generation increases, quicker characterization methods are required. The traditional approach is time-consuming as it has to consider the relevant properties of each combination. To tackle this issue, the researchers came up with a customized microscope that is competent enough to click pictures at nine discrete wavelengths and use these data to calculate the light absorption technology of each sample at each wavelength. The relationship between the optical absorption and wavelength has been considered the identifying fingerprint of the samples.
Two different models have been used to analyze the results: The phase Diagram Model and Emergent Property Model.
The phase diagram model is a physics-backed model that assumes thermodynamic equilibrium, limiting the number of coexisting phases. Presuming that the optical properties of a combination of crystalline phases vary linearly with the ratio of each phase, the model generates a set of phases that best fits the optical absorption spectra.
On the other hand, the second model recognizes emergent properties by picking a 3-metal oxide absorption spectra that a linear combination of signals can not describe.
At the end of this systematic approach, out of the 108 3-metal oxide systems, fifty-one of them were found to have interesting properties. However, only one has been reported in the Inorganic Crystal Structure Database. This list consists of nine channels of optical absorption measurements that can be used as an indicator of a unique property and model results giving the most promising compositions.
This research deals with an in-depth analysis of the Cobalt-Tin-Tantalum system. It exhibits tunable frequency, catalytic activity, and stability in strong electrolytes, which is a rare combination of desired properties. The discovery of the new solid solutions was validated by X-Ray diffraction. New materials were resynthesized using the traditional method of Physical Vapor Deposition. Surprisingly, high transparency in composition was noted. The study of the Cobalt-Tin-Tantalum system reinforces the importance of discovering new materials to develop improved technologies, which are critical in industrial applications.
The researchers hope that their work inspires the materials community and more such high-throughput techniques are found. The objective is to provide a rich dataset with plenty of negative results to improve the performance of Machine Learning models.