Skip to main content

A materials informatics approach to the Pauling’s rules and structure-property relationships in low thermal conductivity materials

Periodic Reporting for period 1 - JG-GH-UCLouvain (A materials informatics approach to the Pauling’s rules and structure-property relationships in low thermal conductivity materials)

Reporting period: 2019-10-01 to 2021-09-30

Inorganic materials are essential to a large part of our modern life and are part of the solution for many of the major problems of modern society. For example, thermoelectric materials, by transforming heat to electricity, could potentially increase the energy efficiency of production sites, aircrafts or cars. The design and discovery of materials is strongly dependent on the understanding of structure-property relationships. While structure-property relationships have always been central to materials science, the advent of materials informatics offers new exciting opportunities to discover structure-property relationships.
A very well established and powerful way to describe inorganic crystal structures is through coordination polyhedra. With the help of these coordination polyhedra, some crystals have been rationalized and predicted. Linus Pauling based his famous five rules on the stability of ionic crystals on these coordination polyhedra and their connections. These five rules can be seen as structure-property relationships with stability as the property.
Within the project, we have statistically assessed these five Pauling rules for a set of 5000 oxides for the first time. We have seen that these rules are only of limited predictive power. The rules two to five only work for roughly 13% of all tested oxides. Unfortunately, these rules cannot be used for a fast evaluation of the stability of materials due to their limited predictive power.
In addition to understanding materials based on their coordination environments, there are other possibilities. Another possibility is bonding analysis using the Crystal Orbital Hamilton Population, for which we have developed tools that automatize these calculations and allow for high-throughput calculations within in this project. This allowed us to test a new implementation into a well-known software package and to use this tool in several ab initio high-throughput studies to understand the results. One of these high-throughput studies identified new ferroelectric materials that could be used in low-power storage devices.
Based on the detailed investigation of a thermoelectric material (14-1-11), we have developed a new design principle to arrive at materials with low thermal conductivity.
Beyond chemical heuristics and design principles, machine-learned interatomic potentials also allow material properties to be calculated in an accelerated manner. As an alternative to heuristics, in this project we investigated how we can use these potentials for vibrational properties. A new recipe for constructing databases on which to base these potentials was developed in the process. We were thus able to calculate the phonon properties of several silicon allotropes in good agreement with ab initio calculations. We have also shown that thermal conductivities can be calculated based on this approach which might allow to discover thermoelectric materials with low thermal conductivities in the future.
In the course of the project, we have seen that tools from materials informatics can indeed be very helpful to find and understand new materials for many applications (thermoelectric materials, ferroelectric materials) and that tools based on coordination environments and also bonding analysis can be used for this.
Within the project, we have shown how important chemical heuristics (intuitive strategies within chemistry) and design principles based on coordination environments and bonding analysis are to understand, find and design new materials. Within an opinion, we have pointed out, how easily old chemical heuristics can nowadays be tested, and completely new ones can be developed based on tools from data analysis and machine learning (Trends Chem. 2021, 3, 86).
We have assessed the predictive power of a famous chemical heuristic, the Pauling rules, which connects the structure of the material to the stability, for the first time (Angew. Chem. Int. Ed. 2020, 59, 7569). We have seen that they are only of limited predictive power and they only work well for less than 13% of all tested oxides. Since they are a cornerstone of solid-state chemistry, this was also a very important result for solid-state chemistry, crystallography, and the understanding of crystalline materials. To do so, we have developed automatic tools to determine coordination environments based on crystal structure data (Acta Cryst B 2020, 76, 683).
We have developed automatic tools to perform bonding analysis for crystalline structure based on electronic structure theory calculations. This contributes to the chemical understanding of materials and will simplify the development of new chemical heuristics based on bonding analysis tools. We have already used these tools to test a new implementation in the program Lobster (J. Comput. Chem 2020, 41, 1931) and to understand and rationalize results from ab intio high-throughput searches for new materials that can be used as ferroelectric materials or in photovoltaics (Proc Natl Acad Sci USA 2021, 118, e2026020118). The tools for automated bonding analysis have already been disseminated and a publication in a peer-reviewed journal is planned.
We have developed a new design principle for thermoelectric materials that is based on a new theory on amorphous-like heat conduction based on the well-known thermoelectric material (Mater. Today Phys. 2021, 100344)
To speed up the search for new thermoelectric materials, we have also tested machine-learned interatomic potentials (J. Chem. Phys. 2020, 153, 044104). We have developed new strategies to build databases to learn these potentials. We have arrived at very good agreement of the computed phonon properties based on the potentials with ab initio reference data for several silicon allotropes. We have therefore shown that transferable interatomic potentials can be developed and used to compute phonon properties. We have also shown that thermal conductivities that are very relevant for thermoelectric materials can be computed with these potentials as well.
With our assessment of all five Pauling rules, we have contributed to the field of solid-state chemistry and the understanding of crystalline materials. We have also shown how classical chemical heuristics can nowadays be assessed due to the new methods in data analysis and machine learning.
We have also developed high-throughput bonding analysis tools. This has not been done before and allows to investigate the chemical bonding situation in a large amount of materials in an automatic way for the first time. We have already used these tools to understand results from ab initio high-throughput searches for new ferroelectric materials. These tools will also simplify to develop new, data-driven chemical heuristics that are based on bonding properties in the future. Again, this was a contribution to the field of solid-state chemistry.
Furthermore, we have shown for the first time that phonon properties of a range of silicon allotropes can be computed accurately based on machine-learned interatomic potentials. This is expected to inspire many future studies in this direction. These machine-learned interatomic potentials therefore allow to access the dynamic stability of materials in a much faster way than typical ab initio-based tools. We expect that this will also help in the search for new materials with low thermal conductivity that would be well-suited for thermoelectric materials in the future.
Figure illustrating the 5 Pauling rules that were assessed within this project.