Scientists use machine learning to develop nanoalloy design board
Scientists used machine learning to develop a nanoscale alloy design map that can help predict the matching of metal pairs that can form bimetallic nanoalloys.
These nanoalloys, also called core-shell nanocluster alloys, in which one metal forms the core and another remains on the surface as a shell, are a new frontier in scientists’ quest for new materials and have applications in biomedicine and in other areas.
It is important to know under what conditions core-shell structures form in nanocluster alloys and which metal forms the core, and which remains on the surface as a shell. A number of factors such as cohesive energy difference, atomic radius difference, surface energy difference, and electronegativity of the two atoms can play a role in the preference of the nucleus and shell of the atoms .
The periodic table has 95 metals of different categories ranging from alkaline to alkaline earth, which can potentially form 4465 pairs. It is experimentally impossible to determine how they behave in the formation of nanocluster alloys. But computers can be programmed to predict the behavior of these pairs and more through machine learning. The machine learns to recognize patterns by introducing a number of patterns with well-defined attributes. The more data entered into the computer, the more precise the recognition of unknown data by the computer will be.
However, scientists faced a stumbling block due to the limited number of experimentally synthesized binary nanoclusters with clear identification of the chemical order of constituents and the few core-shell combinations studied theoretically. Machine learning could not be applied with confidence on a small data set of sizes less than or near 100.
Researchers at the SN Bose Center for Basic Sciences, an independent institute of the Department of Science and Technology, have circumvented this problem by calculating the relative surface-core energy on a variety of possible binary combinations of alkali metals, alkaline earth metals, base metals , transition metals and p-block metals to create a large dataset of 903 binary combinations.
In their paper published in the Journal of Physical Chemistry, they investigated key attributes of core-shell morphology using machine learning statistical tool applied to this large dataset. Core-shell structures with lighter metals having lower atomic numbers in the core were classified as type 1, and those with heavier metals in the core were classified as type 2. A number of attributes were constructed to characterize each data point as a whole. The performance of the ML model was compared with existing experimental data and the ML model was found to be reliable.
Having thus established confidence in the ML model, the dominant attributes driving the core-shell model have now been analyzed. It was found that the relative importance of the key factors depends on the combinations of subsets such as alkali-alkaline-earth metal, transition metal-transition metal, etc. It was also found that if the difference in the cohesive energies between the two types of atoms is very small, the nanoclusters constitute a random mixture of the two metals, and if the difference in the cohesive energies is very large, the atoms are separated into a two-sided structure with one face of A atoms and another face of B atoms called Janus structure named after the two-faced Greek god.
Thus, the attempt to connect ML with nanoscience succeeded in tracing the mixing patterns of metal atoms in nanoclusters and formed a basis for the design map, which can help select metal pairs for nanocluster alloys. This design board developed by the scientists will be tested at the nano-laboratories of Moscow State University as well as at the SN Bose Center.
Another area of study for the SN Bose team was the heterogeneous structure formed at the junction of two dissimilar semiconductors. They established that by using machine learning, the types of hetero-structure used in the hetero-junctions of two semiconductors that are at the heart of devices such as LEDs, solar cells and photovoltaic devices, can be predicted quite accurately.
The ML model designed by the SN Bose team predicted 872 unknown type-2 semiconductor heterostructures where electrons and holes align in semiconductor A and semiconductor B, respectively, giving rise to a hetero- desirable structure for semiconductor gadgets.
The SN Bose Center used machine learning to research cheaper substitutes for natural rare earths. Rare earth compounds with permanent magnetic properties are used in loudspeakers and computer hard drives. Of these, 17 elements of the periodic table like neodymium, lanthanum, etc., are found sparsely on the earth’s crust, and their supply is monopolized by the countries where their mines are located. By painstakingly creating a database of rare earth compounds and their attributes, then building a machine learning model, they predicted a list of potential candidates for permanent magnets that will cost less than $100 per kg.
This work as part of the “National Supercomputing Mission” has added a whole new dynamic to humanity’s quest for new materials.
(With GDP entries)