With the help of AI, materials researchers want to orientate themselves more quickly in the jungle of potential materials. The search for an alternative to lithium in batteries was particularly quick recently. But there is a catch.
Lithium is processed in a brine basin in the Chilean Atacama Desert – here is a bird’s eye view in 2023.
Rodrigo Abd / AP
It is the dream of battery researchers: after a short search, to come across a brand new material that makes electric car batteries cheap, compact and environmentally friendly at the same time – and still covers a thousand kilometers per charge.
Such a search, seeking an alternative to currently used lithium, is important for several reasons. Because battery manufacturers want to become more independent of the raw material. The element threatens to become scarce as electric cars and home storage systems for solar energy, which require lithium batteries, become more widespread.
The mining of lithium is also criticized because it lowers the groundwater level in Chile, for example, or destroys landscapes in open-cast mining.
One of the largest sources of lithium is the mine of the Chilean Sociedad Química y Minera in the Atacama salt flat (photo from January 10, 2013).
Ivan Alvarado / Reuters
At Rincon Mining’s lithium pilot plant, located in Argentina’s Salar del Rincón salt desert, an employee reaches into a container of lithium carbonate.
Agustin Marcarian / Reuters
What is needed are new, more readily available and environmentally friendly materials – for example sodium. But this opens up a wide field: chemical elements can be combined in millions of variations to create new battery materials, comparable to a Lego construction set with many different parts. Finding a variant that combines all the desired features is like looking for a needle in a haystack, which takes many years.
Artificial intelligence shortens the search to a week
Researchers from Microsoft and the Pacific Northwest National Laboratory (PNNL) in the American state of Washington are now showing that artificial intelligence (AI) can be used much faster than conventional methods. Chi Chen’s team from Microsoft found a replacement for a lithium-containing battery component that uses 70 percent less of the metal.
“We condensed twenty years of searching into one week,” says Nathan Baker from Microsoft. The team describes the process on the preprint server Arxiv. The work has not yet been peer-reviewed.
Researchers have been using computers to search for new materials for decades. They simulate new substances to predict their properties. Only materials that have shown the desired properties in the simulation are actually synthesized in the laboratory.
But the usual search method requires a lot of computing time because it has to solve highly complex equations in quantum physics. AI works completely differently. She does not know the rules of physics, but acquires something that resembles human experiential knowledge.
The AI learns connections from materials science
In a learning phase, the AI receives data from thousands of materials that describe both their structure from chemical elements and their properties, i.e. whether the material is stable or how well it conducts electrically charged atoms (ions). From this, the AI learns statistical relationships between the structure and properties of the materials.
The amazing thing is that the AI learns relationships that apply to the entire class of materials, not just to the materials in the training data set. Using the learned model, she can quickly estimate whether a new candidate has the desired characteristics.
Some battery anodes are made of copper foils coated with thermally vaporized lithium metal – photographed here on March 22, 2023 at the CSEM Battery Innovation Hub in Neuchâtel.
Christian Beutler / Keystone
Chi Chen’s team was looking for a material for a solid electrolyte. This component of the battery transports ions between the terminals of the battery. In order to make batteries fireproof, manufacturers want to replace liquid electrolytes with solids. This would enable more compact batteries, for example for small electric aircraft. This solid should also contain as little lithium as possible.
Within 80 hours, the American researchers’ AI selected eighteen candidates from 32 million. This was done with the help of several AI models, each of which filtered for a desired property one after the other. The first tested whether a material was stable. About half a million of the candidates passed. The other models filtered substances according to electronic properties. For example, electrically insulating materials were ruled out. On the other hand, high conductivity for ions is particularly important.
The candidates are examined in the laboratory
Computers cannot guarantee that materials will actually work as predicted. AI can’t do that either. Therefore, the remaining candidates must be synthesized and examined in the laboratory. The PNNL researchers have already examined four of them. They particularly liked one. “The electrolyte contains 70 percent less lithium than other industry announcements,” explains Nathan Baker. A large proportion of the lithium atoms in its crystal lattice are replaced by sodium atoms.
Sodium is considered a cheap and readily available replacement for lithium in battery research because it has similar chemical properties. There is a thousand times more sodium in the earth’s crust than lithium, and there are also large reserves in Europe. The element is obtained from table salt, which is mined in salt mines or from sea water.
The newly found solid-state electrolyte fits well with these efforts. Not only because it contains less lithium itself, which is an important intermediate step towards completely eliminating the element. But also because it conducts both lithium and sodium ions very well. It could therefore be suitable for batteries whose positive poles contain one of the two elements or a mixture of them. So it offers freedom for different battery designs. Chi Chen’s team has built battery prototypes with it, which they are currently testing.
AI can also be used to search for materials for solar cells
PNNL’s materials researchers are not the first to use AI as a discovery tool. Materials scientists have been developing similar AI models for around ten years. Pascal Friederich is looking for new materials, for example for solar cells, that are both highly efficient and durable – something that is not yet possible with today’s materials.
“With AI, we can predict the properties of 100,000 potential materials within a few hours,” explains the expert in computer-aided material development at the Karlsruhe Institute of Technology. “The classic simulation methods take several weeks for this.” So the AI-supported method is more than 100 times faster. However, such comparisons are difficult because they depend heavily on the computing power used, emphasizes Friederich.
AI mostly uses graphics processing units (GPU), while the classic method uses normal processors. With AI currently booming, GPUs are very expensive and not every institution can afford them. This is an opportunity for Microsoft to offer material search using AI as a service in its cloud.
Nathan Baker from Microsoft even speaks of a “new way of doing chemistry”. So will chemists soon be replaced by AI? Pascal Friederich disagrees: “I see AI as an additional tool alongside experiments and classic physical simulations.” The physicist emphasizes that this is the only way to accelerate research.
Some supposed successes are not successes at all
AI alone does not provide new knowledge. The connections between the structure of a material and its properties that it learns are statistical, they are correlations. “We researchers have to use our classic methods to check that the correlations actually correspond to physical connections,” says Friederich. It happens, for example, that a material predicted to be stable cannot be synthesized in the laboratory.
In order for the AI to make fewer mistakes, which cost unnecessary time in the laboratory, it must be able to learn from its mistakes. The test in the laboratory provides new data with which the AI can be retrained in order to become more accurate. “This feedback is currently being intensively researched,” says Friederich.
But even if this succeeds, the AI will continue to make incorrect predictions. The chemistry of solids is complex. During synthesis in the laboratory, other substances can be created depending on the pressure or solvent – not every variant will have the desired properties. In order to filter out the few “gold nuggets”, a lot of laboratory work will also be necessary in the future. But researchers are already working on automating laboratory experiments using robots and AI.
AI is certainly becoming part of the everyday lives of materials scientists. Whether she can actually find materials for an all-round worry-free electric car battery remains to be seen.
Even with AI help, it will probably take some time before the mining of lithium can be dispensed with, as is the case here in the Atacama salt flats.
Ivan Alvarado / Reuters