There is a need for new materials that can help with environmental sustainability.
Nearly 70% of the energy produced globally is wasted as heat, which escapes from industrial and commercial premises such as factories, datacenters, and supermarkets, as well as from engines and machines. Thermoelectric materials could harness this wasted heat and turn it into electricity, but existing materials with this capability have major limitations, such as not being efficient enough, not being available in large quantities, or simply being toxic.
"There's still the quest to discover new materials that have the right thermoelectric performance to make them commercially viable on a large scale," says Keith Butler, a senior lecturer in green energy at the U.K.'s Queen Mary University of London.
However, coming up with novel materials using traditional methods is expensive and can take decades. Part of the problem is that the number of potential materials, created by combining different elements in the periodic table, is immense. Simply looking at inorganic materials made up of four elements, for example, there are thought to be 1010 (10 billion) possible combinations after chemically implausible compositions are ruled out.
Machine learning increasingly is being used to aid material discovery, since it can drastically speed up the process. Especially when it comes to identifying materials that could help tackle the climate crisis, time is of the essence. "Our aim is to make material development at least 10 times faster than the traditional way," says Patrick Teyssonneyre, CEO and co-founder of Singapore-based Xinterra, a start-up that has pioneered a new artificial intelligence (AI)-based platform to help develop novel sustainable materials.
Butler and his colleagues are using AI to search for materials that could help with clean energy, such as materials that could make photovoltaic solar cells more efficient, and new battery materials that are not scarce like lithium and others currently used. In recent work, they used deep learning to predict new thermoelectric materials. To train their model, they used a freely available dataset of about 48,000 inorganic materials whose thermoelectric properties had been calculated by other researchers using quantum mechanics. "The nice thing about using deep learning is that you need to do a minimal amount of engineering of your data before you put it into the model," says Butler.
To search for new thermoelectrics, the model was fed over 54,000 unique material compositions that had unknown thermoelectric properties. They were obtained from The Materials Project, a database with information about known and unknown materials that can be used by researchers trying to invent new substances. Butler and his team found the model performed reasonably well when trained solely with a material's composition, and even better when its structure was also available. They selected 23 materials generated by the model that had promising thermoelectric properties for further analysis. "We take the best performing ones and do some rigorous quantum mechanics calculations on them," says Butler. "It's about reducing the search space as you increase the expense and the accuracy of the calculations."
Their next step will be to try to find synthetic chemists who are interested in making these potential new materials and measuring their thermoelectric properties. When using machine learning models, one of the biggest challenges is that they can rarely predict whether the novel materials they are coming up with can actually be made. Further analysis is needed, such as using other machine learning methods to create maps of existing materials based on their structures and properties; a potential new material can then be compared to these maps to see whether, and where, it fits in. Existing materials are typically clustered together so if a new substance sits close by, it means it is similar and there is a good chance it is synthesizable. "Whether we can actually make these materials or not is absolutely critical to address," says Butler. "Otherwise, (synthetic chemists) will lose confidence in our predictions quite quickly."
Synthesizability is also a concern for Teyssonneyre and his colleagues, who have already made dozens of materials for customers, such as a coating that can be applied to metallic surfaces to insulate it from high temperatures. To help gauge whether a material can actually be made, they can add constraints to their machine learning algorithms, such as solubility requirements and equipment capabilities. Even with such additions, however, it's not always enough. "We also work closely with experts to further filter out designs that might not be possible to make," says Teyssonneyre.
Butler and his team has developed a set of machine learning algorithms specifically for material science applications that they can choose from depending on the problem they are trying to solve. Whereas some of them involve deep learning, which requires a lot of data for accurate predictions, they also can use other machine learning models if data is sparce. A model is typically trained with data provided by clients who are aiming to develop a material for a specific purpose. It can be constrained based on a customer's requirements, such as the incompatibility of certain ingredients or if there is a cost cap.
The approach developed by Teysonneyre and his colleagues also involves generating new data. Machine learning is combined with high-throughput experimentation — a process that uses automated systems to carry out experiments quickly and gain new quality data that can be fed back to the machine learning algorithms to improve their performance.
For example, the team is currently focused on creating unique materials that can be integrated into textiles to absorb carbon dioxide from the air. Once the team's algorithms have predicted several potential new materials, the robotic experimentation system can measure how they perform when incorporated into samples of fabrics placed in vials. "In general, institutions or companies who do these experiments can measure one sample per day," says Teyssonneyre. "In this case, we can measure 50 per day (with our automated experiments)."
Teyssonneyre and his colleagues initially will target the fashion industry with their carbon-capture materials. Clothing is responsible for huge amounts of pollution, both during production and use, which has put the sector under pressure to rethink its practices. "We believe that we can develop a solution that will help this industry to overcome some of their environmental challenges," says Teyssonneyre. "Our target is to come up with a prototype of a carbon-capture piece of fabric by the end of March."
Sandrine Ceurstemont is a freelance science writer based in London, U.K.