A team of researchers from Stanford and MIT used AI to predict the performance over the life of lithium-ion battery cells.
Normally, AI would require data from when a battery has started to fail to be able to predict its future performance. This data might not be available until the battery has been cycled enough times that it takes several months. The AI that the researchers developed was able to predict the batterys lifetime performance in hours.
This was even though the battery was still at its maximum. William Chueh (a Stanford materials scientist and one of the leading authors of the 2019 paper) says, "Prior to our work, nobody thought it was possible." Chueh and his collaborators did it again earlier in the year.
Chueh and his collaborators described an experiment where an AI was capable of finding the best method to fast-charge a lithium-ion battery for a 10-minute period.
Experts believe fast-charging batteries are crucial for electric vehicle adoption. However, dumping enough energy to charge a cell in the time it takes to fill up an empty tank of gas can quickly destroy its performance.
Finding the right balance between battery life and charge speed is key to fast-charging batteries. Chueh says that there are literally infinite ways to charge a battery. He compares it with trying to find the best way to pour water into an empty bucket.
It can be difficult to choose the right one, and it is time-consuming. AI can help.
Chueh and his team were able to create a fast-charging protocol to charge a lithium-ion battery within a month. Without the help of AI, it would normally take two years to achieve these same results.
Chueh says, "At the end the day we see our job is to accelerate the pace of battery research and development." Its time-consuming, whether its finding new chemistry or making safer batteries. We are trying to save time."
Read Blog: AI to play a greater role in supporting high-quality development
The performance of batteries has increased dramatically over the past decade, while their costs have dropped. This is good news, as many experts believe that electrification of all energy systems will be key to decarbonizing the planet.
Researchers like Chueh are concerned that battery innovation is not moving fast enough. It is because batteries are so complex. A better battery requires constant optimization at each stage of the manufacturing process.
It all comes down to using cheaper raw materials, better chemicals, and more efficient manufacturing methods. There are many parameters that can be optimized. Of course, again in one area (e.g. energy density) will often come at the expense of gains in another.
The problem AI was created to solve is finding optimal solutions in large search spaces. However, until recently, data was a problem for battery-building AIs.
"Historically, battery data was very difficult to obtain because its never shared between researchers or companies," says Bruis Van Vlijmen, a Stanford data scientist who works on battery analytics. "There is a high degree of secrecy and proprietary information." After their 2019 paper, Chueh with his colleagues made all their battery data public so that other researchers could use it to train their AI algorithms.
It was the largest ever collection of battery performance data.
The lack of quality data is a problem that Ian Foster, director of the data science division at Argonne National Laboratory, has been dealing with for many years.
Foster and his lab colleagues have been working together to build a database of molecules. This database can be used by a machine-learning algorithm to search for chemicals that could improve the performance of a batterys electrodes.
Electrolyte chemistries, like all elements of a battery, can be modified to improve desirable properties such as energy density or decrease its toxicity. Foster says that the process of finding new electrolyte materials was historically a trial-and-error one. "Our goal is AI to explore the virtually infinite number of materials."
The Argonne team published a pair of papers in late 2019 that described how they used a database of 133,000 organic molecules as well as the supercomputer at the lab to create precise simulations of the properties of these molecules.
These molecules can have up to nine "heavy" atoms. The idea was to use the database to train a machine-learning algorithm to find desirable properties in small datasets so that it could explore larger databases of possible materials.
There are many ways these atoms could be combined, so most battery electrolytes can have up to 20 heavy atoms. Another database that contains organic molecules with up to 17 heavy elements is 166 billion. This would make it difficult for an AI to find promising candidates without knowing what it was looking at.
Foster said that it is still early days for Argonne's electrolyte hunt algorithm. Although it has not yet identified any new materials, the next step in the process will be to build a physical cell with that electrolyte material for experimentation.
These data can be used to refine the algorithm and narrow down its selection to better candidates. Foster says that the process of going from a large number of electrolytes to one that will actually be used in millions of cars takes a long time.
Machine learning is designed to speed up the experimentation process.
Read Blog: 8 Forthcoming Trends in Artificial Intelligence and Deep Learning
Fosters team has been working with scientists from a dozen research institutes and companies to facilitate the sharing of statistics across organizations.
They plan to use Data Station, a platform created at the University of Chicago that allows researchers to train machine-learning models using a pool of information from different groups. This is without allowing outsiders to have direct access to the data. The platform allows researchers to upload a machine learning model, train it on the data, then return it to them.
Scientists dont have access to the data but can see if the model was able to predict battery prices. Foster and his colleagues hope that this will allay peoples fears of losing their proprietary data to others while still allowing for the creation of massive data sets.
Even without large shared databases, AI is already being used in battery development. According to a paper published in Frontiers of Energy Research this summer, AI has been used in an astonishing number of battery research applications within the last year.
It has been used to study the stability of lithium metal anodes using molecules. This can dramatically increase energy density, but also pose safety risks. Machine learning was used to identify potential cathode coatings that could improve the performance and safety of batteries with solid electrolytes.
These are safer than the liquid electrolytes in current batteries. AI was also used by researchers to enhance their understanding of existing batteries. This included optimizing battery management systems and creating mathematical models of batteries that can simulate their performance in EVs.
A book was even written by an AI summarizing the current research on lithium-ion battery technology.
Alpha Lee, a University of Cambridge statistical physicist, says that there is still a lot of potential in existing batteries materials.
He suggests that better software can be used to "program" the battery. His recent research uses machine learning to discover new predictors of the health of the battery. "Innovations made in battery software will reap the benefits of the digital revolutions scalability and help usher in a new age of energy storage technologies.
Next, machine learning techniques will be used to create batteries for our cars and gadgets. InoBat, a Slovakian-based company that was founded in 2018, could be the leader.
Wildcat Discovery Technologies in California developed an AI-powered platform that allows the company to quickly prototype new battery chemistries and make custom cells for electric cars. Marian Bocek (CEO of InoBat) said that the AI platform allows for a comprehensive exploration of new lithium-ion chemicals, which could dramatically accelerate the discovery process.
The AI platform can simulate the performance of a battery when multiple variables are changed at once, instead of tweaking each component one at a time, and then exhaustively testing them all.
Bocek says that the road to discovering new cell chemistry takes 10 times longer than a traditional laboratory. He also compares InoBat's AI-fueled research with the use of automated drug discovery within the pharmaceutical industry.
We are moving away from the one size fits most approach that dominates the EV industry.
InoBats first intelligent battery, which was designed using AI, was unveiled last week by the company. Bocek stated that the battery would increase the range of a "best-in-class" EV by almost 20 percent during the announcement.
However, you wont find it in an average EVs battery pack anytime soon. InoBat is not like the major lithium-ion cell producers, such as Samsung or Panasonic, but it is a boutique battery manufacturer.
InoBat focuses on specialized vehicles, such as high-performance EVs and electric aircraft. It can also do low-volume production to create cells that are tailored to customer needs. Bocek says that the company is the only one like it in the market capable of developing a custom solution in terms of cell format and energy density.
Bocek claims that the companys pilot plant will begin producing batteries by the end of next year. The plant will initially produce only 100 megawatt-hours per year of AI-designed batteries.
This is roughly half of the annual production volume at Teslas Gigafactory, Nevada. Bocek said that the company plans to increase its production to a 10-gigawatt hour facility in five years. This will allow it to match the planned output of Teslas pilot plant in California, which company officials announced last month at the Battery Day event.